Lidl Introduces the European Payment Method Wero in Germany
5 by doener | 2 comments on Hacker News.
Sunday, July 26, 2026
Saturday, July 25, 2026
New top story on Hacker News: Show HN: I made some transistor animations
Show HN: I made some transistor animations
30 by stunningllama | 2 comments on Hacker News.
Hi HN, I made some animations of the most important kinds of transistors using my semiconductor simulation, details of which are on the page. I tried to make the visuals as realistic as possible while also aiming for clarity. If you want to go beyond the charge carriers and look at, for example, the electric field, you can do so in the simulation software. The desktop software also has less common devices like IBGTs and SCRs that have similar animations. The last thread about my software was posted here about a year ago: https://ift.tt/izokYd3
30 by stunningllama | 2 comments on Hacker News.
Hi HN, I made some animations of the most important kinds of transistors using my semiconductor simulation, details of which are on the page. I tried to make the visuals as realistic as possible while also aiming for clarity. If you want to go beyond the charge carriers and look at, for example, the electric field, you can do so in the simulation software. The desktop software also has less common devices like IBGTs and SCRs that have similar animations. The last thread about my software was posted here about a year ago: https://ift.tt/izokYd3
Friday, July 24, 2026
New top story on Hacker News: Nvidia, Microsoft, Meta warn against overregulating open-weight models
Nvidia, Microsoft, Meta warn against overregulating open-weight models
55 by louiereederson | 79 comments on Hacker News.
Letter: https://ift.tt/5Jiz09D... [pdf] https://ift.tt/sRLG0CZ , https://ift.tt/ouUYjDg https://ift.tt/aL9hbDN... , https://ift.tt/IQfKvze
55 by louiereederson | 79 comments on Hacker News.
Letter: https://ift.tt/5Jiz09D... [pdf] https://ift.tt/sRLG0CZ , https://ift.tt/ouUYjDg https://ift.tt/aL9hbDN... , https://ift.tt/IQfKvze
Thursday, July 23, 2026
New top story on Hacker News: Show HN: Trifle – Open-source analytics that stores answers, not events
Show HN: Trifle – Open-source analytics that stores answers, not events
7 by iluzone | 0 comments on Hacker News.
Trifle is an open-source time-series analytics library that aggregates nested counters instead of storing raw events. All in the database you already have. After rebuilding it twice over 10 years, it now tracks ~1B events a day at my day job. It started in 2015 as my own Rails APM. I plugged into ActiveSupport::Notifications, got a few small users, and one bigger one whose scraping app broke everything. That sparked the core idea: aggregate counters into pre-defined time buckets, so a single write increments multiple buckets at once. The APM eventually faded away without much traction. Later in 2021 I needed analytics at my day job. Instead of going for something out there I revised the idea of Trifle as a more generic analytics library, borrowing some data warehouse ideas. First used Redis, then Postgres, eventually MongoDB. Hence why Trifle::Stats comes with multiple drivers that keep the DSL unified while storage layer changes with your needs. In our case (huge write volume, some reads) PG read faster but slowed on large writes. The nested values are the whole trick here. Single: Trifle::Stats.track( key: 'requests::aws::s3_uploads', values: { count: 1, status: { request.response_code => 1 }, size: payload.bytes, duration: { sum: request.duration, count: 1 } } ) builds up counts for requests, success rate, result status codes, duration for multiple time buckets at once. Single bucket from 2am then looks like: { count: 14, status: { 200: 12, 500: 2 }, size: 5628341, duration: { sum: 43, count: 14 } } If request.duration is in seconds, then sum stored under duration would be in seconds as well. Success rate is never stored, but it is calculated by dividing 200s over total number of requests. Same with average duration: sum over count. You ask for a metrics key, granularity and timeframe and you get back aggregated values at each point. Ready for charts or to answer "Average response time over last 30 days". There's a Series wrapper for aggregating and formatting values for charts in a simple call. And as building dashboards is not as much fun for other devs as I thought, I built Trifle App - a visual layer with dashboards, scheduled digests and alerts. It's written in Elixir, so I ported the library to Elixir too. And later to Go for a CLI. All three are compatible, write in one and read in another. Today we track activity from over 100M background jobs a day which turns into about 1B events. It runs surprisingly cheap when you're willing to trade some safety away (turn off journaling and write concerns in Mongo). 3-node Hetzner MongoDB cluster where the primary does 20% utilization costs us around $1k/month. It has its limitations. Payloads can't hold tens of thousands of keys. Documents becomes too large to update efficiently. Some planning ahead is needed. And then there are no dimensions. Sometimes you can nest them (country - there are only so many countries), sometimes it's better to have dedicated metrics key per dimension (customer - growing forever). That multiplies tracked events, hence 1B events from 100M jobs. The libraries are MIT. The App is source-available under ELv2 - free to self-host and paid cloud if you want it managed. I build this on the side with no investor money to burn on a free service. Happy to answer anything about architecture, storage models, my failures or why I didn't give up on this yet.
7 by iluzone | 0 comments on Hacker News.
Trifle is an open-source time-series analytics library that aggregates nested counters instead of storing raw events. All in the database you already have. After rebuilding it twice over 10 years, it now tracks ~1B events a day at my day job. It started in 2015 as my own Rails APM. I plugged into ActiveSupport::Notifications, got a few small users, and one bigger one whose scraping app broke everything. That sparked the core idea: aggregate counters into pre-defined time buckets, so a single write increments multiple buckets at once. The APM eventually faded away without much traction. Later in 2021 I needed analytics at my day job. Instead of going for something out there I revised the idea of Trifle as a more generic analytics library, borrowing some data warehouse ideas. First used Redis, then Postgres, eventually MongoDB. Hence why Trifle::Stats comes with multiple drivers that keep the DSL unified while storage layer changes with your needs. In our case (huge write volume, some reads) PG read faster but slowed on large writes. The nested values are the whole trick here. Single: Trifle::Stats.track( key: 'requests::aws::s3_uploads', values: { count: 1, status: { request.response_code => 1 }, size: payload.bytes, duration: { sum: request.duration, count: 1 } } ) builds up counts for requests, success rate, result status codes, duration for multiple time buckets at once. Single bucket from 2am then looks like: { count: 14, status: { 200: 12, 500: 2 }, size: 5628341, duration: { sum: 43, count: 14 } } If request.duration is in seconds, then sum stored under duration would be in seconds as well. Success rate is never stored, but it is calculated by dividing 200s over total number of requests. Same with average duration: sum over count. You ask for a metrics key, granularity and timeframe and you get back aggregated values at each point. Ready for charts or to answer "Average response time over last 30 days". There's a Series wrapper for aggregating and formatting values for charts in a simple call. And as building dashboards is not as much fun for other devs as I thought, I built Trifle App - a visual layer with dashboards, scheduled digests and alerts. It's written in Elixir, so I ported the library to Elixir too. And later to Go for a CLI. All three are compatible, write in one and read in another. Today we track activity from over 100M background jobs a day which turns into about 1B events. It runs surprisingly cheap when you're willing to trade some safety away (turn off journaling and write concerns in Mongo). 3-node Hetzner MongoDB cluster where the primary does 20% utilization costs us around $1k/month. It has its limitations. Payloads can't hold tens of thousands of keys. Documents becomes too large to update efficiently. Some planning ahead is needed. And then there are no dimensions. Sometimes you can nest them (country - there are only so many countries), sometimes it's better to have dedicated metrics key per dimension (customer - growing forever). That multiplies tracked events, hence 1B events from 100M jobs. The libraries are MIT. The App is source-available under ELv2 - free to self-host and paid cloud if you want it managed. I build this on the side with no investor money to burn on a free service. Happy to answer anything about architecture, storage models, my failures or why I didn't give up on this yet.
Wednesday, July 22, 2026
Tuesday, July 21, 2026
Monday, July 20, 2026
Sunday, July 19, 2026
New top story on Hacker News: Show HN: I replaced a $120k bowling center system with $1,600 in ESP32s
Show HN: I replaced a $120k bowling center system with $1,600 in ESP32s
123 by section33 | 10 comments on Hacker News.
I might be the only SRE on Earth with his own bowling center. It's a more in-depth gig than you'd think. My family and I bought an abandoned 8-lane bowling center in the rural mid-west. In our small town there weren't many recreation options for families. You've heard of a food desert? This is an R&R desert. It had been abandoned for a good reason. The roof leaks, the electrical system was constantly surging, and my 70-year-old bowling equipment (still) doesn't work perfectly. The system that keeps your score is particularly interesting to me. It's the thing you watch during your game, but it fades into the background beyond that. Turns out these things are really cool, but absurdly expensive. Ours was installed in 2008 and cost six figures. It's calculating ball speed and trajectory, camera-based pin detection (object detection and trig, on ICs!), runs the fouling, the animations, the pinsetting machine and ball return. Very cool stuff for its age. From the business perspective, my facility only cost me $105k. To forklift-replace the score keeping system runs anywhere between $80-$120k, depending on features, vendor, and unit age. No upgrades or service contracts, mind you, and every feature and customization is a new line item. That's for a 1:1 replacement on a system installed in 2008. Incredible, given how fast the tech world moves. Replacement parts cost a shocking $4000 per pair of lanes. But wait, the bowling machines themselves are 70 years old, so what's this "advanced" system actually doing back there? Actuating a single relay to trigger that big old machine. Everything else is strictly mechanical. Meanwhile I've got a six-figure invoice in my hand. I'm upset. Given the state of open hardware, computer vision, real-time event streaming, and open source running megascale products worldwide, there had to be a way to do this myself. So far I've built an equivalent prototype for about $200 per lane-pair, $400 if you're fancy. ESP32 and ESPNow with an RS485 fallback, reporting to a raspberry pi lane computer that's really just redis and a state machine bolted to an ESP32 gateway for the mesh. Since it's all ESP32, I've got a fistful of spare controllers in a drawer, pre-flashed or waiting to be. All common off-the-shelf hardware: microcontrollers wired to relays, optocouplers, and IR-break-beam sensors, each running slightly different firmware. Writing the firmware and protocol is the actual hard part. It's an ESPNow star-topology mesh: each node emits events from its sensors and accepts commands for its controls, reporting to a gateway node connected to the raspi over UART. From there it's event streaming: RX packets get translated and tossed into redis, commands relay back out to the mesh as needed. RS485 sits underneath as a wired fallback for noisy RF environments. Once the data's in redis, it's familiar middleware/React/websocket/pub-sub stuff. Any React dev can build their own UI and bowling animations. Since it all runs on commodity hardware, I can do legit anything I want as the proprietor, and I own all my data. Repairs take five minutes; I can swap the rig on a lane pair in under 10. I'd bet a house like mine could go from zero to running in an hour or two. We're calling it OpenLaneLink, and I plan to open source the hardware, firmware, and software stack when it's ready. Bowling is fun, and I want to help keep it affordable for alleys like mine. I hate vendor lock-in. I'm not a fan of closed systems, calling support for every hiccup, or paying to "white label" my own equipment. Want to go Tron-themed for a night? Good luck finding a neon neumorphic theme in something bought at the turn of the century. All that bugged me. Sure, bowling equipment is niche, but the open hardware and software landscape is amazing. Thanks for reading! Let me know if anyone's interested in more posts about this bowling nonsense.
123 by section33 | 10 comments on Hacker News.
I might be the only SRE on Earth with his own bowling center. It's a more in-depth gig than you'd think. My family and I bought an abandoned 8-lane bowling center in the rural mid-west. In our small town there weren't many recreation options for families. You've heard of a food desert? This is an R&R desert. It had been abandoned for a good reason. The roof leaks, the electrical system was constantly surging, and my 70-year-old bowling equipment (still) doesn't work perfectly. The system that keeps your score is particularly interesting to me. It's the thing you watch during your game, but it fades into the background beyond that. Turns out these things are really cool, but absurdly expensive. Ours was installed in 2008 and cost six figures. It's calculating ball speed and trajectory, camera-based pin detection (object detection and trig, on ICs!), runs the fouling, the animations, the pinsetting machine and ball return. Very cool stuff for its age. From the business perspective, my facility only cost me $105k. To forklift-replace the score keeping system runs anywhere between $80-$120k, depending on features, vendor, and unit age. No upgrades or service contracts, mind you, and every feature and customization is a new line item. That's for a 1:1 replacement on a system installed in 2008. Incredible, given how fast the tech world moves. Replacement parts cost a shocking $4000 per pair of lanes. But wait, the bowling machines themselves are 70 years old, so what's this "advanced" system actually doing back there? Actuating a single relay to trigger that big old machine. Everything else is strictly mechanical. Meanwhile I've got a six-figure invoice in my hand. I'm upset. Given the state of open hardware, computer vision, real-time event streaming, and open source running megascale products worldwide, there had to be a way to do this myself. So far I've built an equivalent prototype for about $200 per lane-pair, $400 if you're fancy. ESP32 and ESPNow with an RS485 fallback, reporting to a raspberry pi lane computer that's really just redis and a state machine bolted to an ESP32 gateway for the mesh. Since it's all ESP32, I've got a fistful of spare controllers in a drawer, pre-flashed or waiting to be. All common off-the-shelf hardware: microcontrollers wired to relays, optocouplers, and IR-break-beam sensors, each running slightly different firmware. Writing the firmware and protocol is the actual hard part. It's an ESPNow star-topology mesh: each node emits events from its sensors and accepts commands for its controls, reporting to a gateway node connected to the raspi over UART. From there it's event streaming: RX packets get translated and tossed into redis, commands relay back out to the mesh as needed. RS485 sits underneath as a wired fallback for noisy RF environments. Once the data's in redis, it's familiar middleware/React/websocket/pub-sub stuff. Any React dev can build their own UI and bowling animations. Since it all runs on commodity hardware, I can do legit anything I want as the proprietor, and I own all my data. Repairs take five minutes; I can swap the rig on a lane pair in under 10. I'd bet a house like mine could go from zero to running in an hour or two. We're calling it OpenLaneLink, and I plan to open source the hardware, firmware, and software stack when it's ready. Bowling is fun, and I want to help keep it affordable for alleys like mine. I hate vendor lock-in. I'm not a fan of closed systems, calling support for every hiccup, or paying to "white label" my own equipment. Want to go Tron-themed for a night? Good luck finding a neon neumorphic theme in something bought at the turn of the century. All that bugged me. Sure, bowling equipment is niche, but the open hardware and software landscape is amazing. Thanks for reading! Let me know if anyone's interested in more posts about this bowling nonsense.
Saturday, July 18, 2026
Friday, July 17, 2026
Thursday, July 16, 2026
Wednesday, July 15, 2026
Tuesday, July 14, 2026
Monday, July 13, 2026
Sunday, July 12, 2026
Saturday, July 11, 2026
New top story on Hacker News: The Oral History of the Dinosaur Input Device (DID)
The Oral History of the Dinosaur Input Device (DID)
3 by fabiensanglard | 0 comments on Hacker News.
3 by fabiensanglard | 0 comments on Hacker News.
Friday, July 10, 2026
Thursday, July 9, 2026
New top story on Hacker News: Show HN: Abralo – Free, easy way to run several Claude Code agents in one window
Show HN: Abralo – Free, easy way to run several Claude Code agents in one window
15 by cwbuilds | 4 comments on Hacker News.
Hi guys, I've been using Claude Code for almost everything lately. Have given one an email account so it can research business leads, draft emails, fact-check them and clear them with me before sending (works really well by the way). I also tend to have a few Claude Code agents running at any one time for coding. I used to create a split terminal to manage them from there, but found working in the terminal all day pretty depressing and, more importantly, found it hard to follow Claude Code's process and see which agents needed my immediate attention. I tried Anthropic's VS Code Claude Code extension and it had a great UI (more info on Claude Code's process and easier to read), but it crashed my PC when I ran more than 3 and I couldn't watch multiple agents in parallel (had to constantly switch between them). So I built a lightweight Tauri desktop app which lets you run multiple Claude Code agents in one window alongside each other. It's easier to read the output and see which agents need your attention than a terminal. Have been using this all day everyday instead of an IDE and have obsessed over every detail to make sure it's easy-to-use, but also lightweight and fast (so you can manage multiple agents without your PC crashing). There are some nice features like better usage alerts for when you're going to hit your 5-hour and weekly limits (with sparklines to show when usage peaked, and which agents are the most token-intensive). It's free to use (you just need to log in with your existing Claude Code account) for up to 4 agents simultaneously. This app doesn't store your Claude Code account details and doesn't store any of your interactions with Claude Code. They remain between you and Anthropic. It's compatible with Windows, MacOS and 64-bit Linux. Would really appreciate any feedback, so if you have any thoughts, issues or suggestions please let me know. Thanks, Chris
15 by cwbuilds | 4 comments on Hacker News.
Hi guys, I've been using Claude Code for almost everything lately. Have given one an email account so it can research business leads, draft emails, fact-check them and clear them with me before sending (works really well by the way). I also tend to have a few Claude Code agents running at any one time for coding. I used to create a split terminal to manage them from there, but found working in the terminal all day pretty depressing and, more importantly, found it hard to follow Claude Code's process and see which agents needed my immediate attention. I tried Anthropic's VS Code Claude Code extension and it had a great UI (more info on Claude Code's process and easier to read), but it crashed my PC when I ran more than 3 and I couldn't watch multiple agents in parallel (had to constantly switch between them). So I built a lightweight Tauri desktop app which lets you run multiple Claude Code agents in one window alongside each other. It's easier to read the output and see which agents need your attention than a terminal. Have been using this all day everyday instead of an IDE and have obsessed over every detail to make sure it's easy-to-use, but also lightweight and fast (so you can manage multiple agents without your PC crashing). There are some nice features like better usage alerts for when you're going to hit your 5-hour and weekly limits (with sparklines to show when usage peaked, and which agents are the most token-intensive). It's free to use (you just need to log in with your existing Claude Code account) for up to 4 agents simultaneously. This app doesn't store your Claude Code account details and doesn't store any of your interactions with Claude Code. They remain between you and Anthropic. It's compatible with Windows, MacOS and 64-bit Linux. Would really appreciate any feedback, so if you have any thoughts, issues or suggestions please let me know. Thanks, Chris
Wednesday, July 8, 2026
Tuesday, July 7, 2026
New top story on Hacker News: Show HN: Halo – open-source, tamper-evident runtime evidence for AI agents
Show HN: Halo – open-source, tamper-evident runtime evidence for AI agents
4 by brian_kuan | 0 comments on Hacker News.
Hi HN, I'm Brian, I spent the last few years at Vanta (YC W18), helping startups and enterprises become compliant and I recently started exploring what that might look like in a post-agentic world. The problem Halo solves is: when a company buys an AI agent from a vendor and gives it access to their data, they have no way to check what the agent did with that data. Vendors may have built observability dashboards and audit logs, but those are editable and partisan. SOC 2 and ISO 27001 audit a company's controls, but controls are less predictive when the software is agentic. TLDR: give an agent the same prompt 50 times, and you get 50 slightly different actions/answers - so the only thing worth auditing in a post-agentic world is what happened at runtime. Halo is an open-source project that produces agent runtime evidence. It's a small recorder that records every action an agent takes (eg. tool calls, model calls, data access, etc), and becomes a record in an append-only log. It's hash-chained, so anyone can re-verify. Run the following command to see a fictional example: uvx --from halo-record halo demo --serve Then, delete a line from one of the .jsonl files and reload, and the report will catch that it's been tampered with. To wire up your own agent, run this line of Python: agent = trace(run_my_agent, profile="my-agent", log="audit.jsonl") Then use this to generate a real report and give it to your customers: halo report audit.jsonl -o report.html Disclaimer: this proves integrity, not completeness (as a self-held chain proves nothing was edited but does NOT prove that nothing was omitted). Catching this requires a witness outside the vendor and is what I'm working on next. Halo is Apache-2.0, contains zero runtime dependencies, and is about 4,300 lines of Python with 125 tests (if you prefer TypeScript, here's that repo: https://ift.tt/DspeUv8 ). Give it a try, and please let me know if you have any feedback!
4 by brian_kuan | 0 comments on Hacker News.
Hi HN, I'm Brian, I spent the last few years at Vanta (YC W18), helping startups and enterprises become compliant and I recently started exploring what that might look like in a post-agentic world. The problem Halo solves is: when a company buys an AI agent from a vendor and gives it access to their data, they have no way to check what the agent did with that data. Vendors may have built observability dashboards and audit logs, but those are editable and partisan. SOC 2 and ISO 27001 audit a company's controls, but controls are less predictive when the software is agentic. TLDR: give an agent the same prompt 50 times, and you get 50 slightly different actions/answers - so the only thing worth auditing in a post-agentic world is what happened at runtime. Halo is an open-source project that produces agent runtime evidence. It's a small recorder that records every action an agent takes (eg. tool calls, model calls, data access, etc), and becomes a record in an append-only log. It's hash-chained, so anyone can re-verify. Run the following command to see a fictional example: uvx --from halo-record halo demo --serve Then, delete a line from one of the .jsonl files and reload, and the report will catch that it's been tampered with. To wire up your own agent, run this line of Python: agent = trace(run_my_agent, profile="my-agent", log="audit.jsonl") Then use this to generate a real report and give it to your customers: halo report audit.jsonl -o report.html Disclaimer: this proves integrity, not completeness (as a self-held chain proves nothing was edited but does NOT prove that nothing was omitted). Catching this requires a witness outside the vendor and is what I'm working on next. Halo is Apache-2.0, contains zero runtime dependencies, and is about 4,300 lines of Python with 125 tests (if you prefer TypeScript, here's that repo: https://ift.tt/DspeUv8 ). Give it a try, and please let me know if you have any feedback!
Monday, July 6, 2026
Sunday, July 5, 2026
Saturday, July 4, 2026
Friday, July 3, 2026
Thursday, July 2, 2026
New top story on Hacker News: The US Government Is Now a Shareholder in 26 Companies
The US Government Is Now a Shareholder in 26 Companies
66 by measurablefunc | 57 comments on Hacker News.
66 by measurablefunc | 57 comments on Hacker News.
Wednesday, July 1, 2026
New top story on Hacker News: Show HN: Morph Reflexes – Multi-head classifiers for agent traces
Show HN: Morph Reflexes – Multi-head classifiers for agent traces
11 by bhaktatejas922 | 1 comments on Hacker News.
The most common failures for production agents are behavioral: looping, reasoning leakage, user frustration, and more. Using a frontier model like GPT or Sonnet to judge every turn is too expensive and slow to run at scale. To solve this, we built Reflexes: semantic signals from agent traces, served fast and cheap over API. Built on custom kernels and a custom inference engine forked from vLLM. Under the hood, it is a small LLM architected around multi-head inference. Small models need to be trained for specific tasks, but running 50 separate small models on the same input for 50 tasks makes no sense. How it works: We use a modern LLM with hybrid attention and remove the decode step. We built an inference engine that lets prefill compute be 99% reused from reflex to reflex, similar in spirit to older 2019-era BERT/HYDRA and older multiple-head techniques. we built the inference engine to reuse the KV/cache across inputs and compute across all reflexes. One shared backbone reads the trace once, then many heads classify different signals. Our inference engine reuses the same KV/cache and compute across all reflexes, giving us sub-30ms inference with less than 0.1% overhead for each additional reflex. We took the same high-level idea and did the hard work to make it work with a modern architecture and attention. On it, we can run inference in under 30ms and serve the full request in under 90ms. If you run 4 reflexes or 100, the extra overhead is less than 2ms. Why does optimizing this matter? If you’re even a medium-sized startup, you’re dealing with tens of thousands of agent runs and millions of turns. If you want to track things like user frustration rates over time, frontier LLM-as-judge does not scale. I built a similar stack at Tesla. When ML engineers needed to sample data across petabytes for signals like `is_camera_obfuscated=true`, along with 200 other things, you need to 1) spin them up quickly 2) run at scale efficiently What it is not: A dashboard. 99% of dashboards go unused. 100% API first and made for devs who want to use this to trigger their own stuff. vibetrain a custom reflex in our dashboard, and/or then let it self improve in production: https://ift.tt/g0OYzTq Docs: https://ift.tt/mxT1osf I’d love feedback from people running agents in prod: what sorts of things do you wish you could track over time across 100% of turns but cant right now? TLDR: semantic signals from agent traces, super fast, cheap via API
11 by bhaktatejas922 | 1 comments on Hacker News.
The most common failures for production agents are behavioral: looping, reasoning leakage, user frustration, and more. Using a frontier model like GPT or Sonnet to judge every turn is too expensive and slow to run at scale. To solve this, we built Reflexes: semantic signals from agent traces, served fast and cheap over API. Built on custom kernels and a custom inference engine forked from vLLM. Under the hood, it is a small LLM architected around multi-head inference. Small models need to be trained for specific tasks, but running 50 separate small models on the same input for 50 tasks makes no sense. How it works: We use a modern LLM with hybrid attention and remove the decode step. We built an inference engine that lets prefill compute be 99% reused from reflex to reflex, similar in spirit to older 2019-era BERT/HYDRA and older multiple-head techniques. we built the inference engine to reuse the KV/cache across inputs and compute across all reflexes. One shared backbone reads the trace once, then many heads classify different signals. Our inference engine reuses the same KV/cache and compute across all reflexes, giving us sub-30ms inference with less than 0.1% overhead for each additional reflex. We took the same high-level idea and did the hard work to make it work with a modern architecture and attention. On it, we can run inference in under 30ms and serve the full request in under 90ms. If you run 4 reflexes or 100, the extra overhead is less than 2ms. Why does optimizing this matter? If you’re even a medium-sized startup, you’re dealing with tens of thousands of agent runs and millions of turns. If you want to track things like user frustration rates over time, frontier LLM-as-judge does not scale. I built a similar stack at Tesla. When ML engineers needed to sample data across petabytes for signals like `is_camera_obfuscated=true`, along with 200 other things, you need to 1) spin them up quickly 2) run at scale efficiently What it is not: A dashboard. 99% of dashboards go unused. 100% API first and made for devs who want to use this to trigger their own stuff. vibetrain a custom reflex in our dashboard, and/or then let it self improve in production: https://ift.tt/g0OYzTq Docs: https://ift.tt/mxT1osf I’d love feedback from people running agents in prod: what sorts of things do you wish you could track over time across 100% of turns but cant right now? TLDR: semantic signals from agent traces, super fast, cheap via API
Tuesday, June 30, 2026
Monday, June 29, 2026
Sunday, June 28, 2026
New top story on Hacker News: I used Claude Code to get a second opinion on my MRI
I used Claude Code to get a second opinion on my MRI
106 by engmarketer | 144 comments on Hacker News.
106 by engmarketer | 144 comments on Hacker News.
Saturday, June 27, 2026
Friday, June 26, 2026
New top story on Hacker News: Ask HN: Is "no source code was copied" still a sufficient copyright defense?
Ask HN: Is "no source code was copied" still a sufficient copyright defense?
17 by oscgam1 | 13 comments on Hacker News.
We are all familiar with the Corgi event: https://ift.tt/aMoqI1B With the barrier to create new apps having dropped significantly thanks to LLMs, I am seeing more cases about copyright and unfair competition. I've seen and participated in some of these cases. Usually expert witnesses are required. Curious to hear the community stance on this one. "Now software developers are feeling what authors and artist felt". https://ift.tt/LbwWgmx There are several claims of: Copying UI is Ok, your product is not undifferentiated enough. Here is a legal assessment of the situation: https://ift.tt/AIoZq85
17 by oscgam1 | 13 comments on Hacker News.
We are all familiar with the Corgi event: https://ift.tt/aMoqI1B With the barrier to create new apps having dropped significantly thanks to LLMs, I am seeing more cases about copyright and unfair competition. I've seen and participated in some of these cases. Usually expert witnesses are required. Curious to hear the community stance on this one. "Now software developers are feeling what authors and artist felt". https://ift.tt/LbwWgmx There are several claims of: Copying UI is Ok, your product is not undifferentiated enough. Here is a legal assessment of the situation: https://ift.tt/AIoZq85
Thursday, June 25, 2026
New top story on Hacker News: Show HN: OpenKnowledge – open source AI-first alternative to Obsidian/Notion
Show HN: OpenKnowledge – open source AI-first alternative to Obsidian/Notion
27 by engomez | 5 comments on Hacker News.
Hi HN, Nick here. We’re launching OpenKnowledge ( https://ift.tt/PN4MyL5 ), a “what you see is what you get” markdown editor that has direct integrations with Claude, Codex, and Cursor. Available as MacOS app or CLI. Fully free/local and OSS ( https://ift.tt/GvhFJlw ). We built this because we wanted a “Google docs” like experience for writing and sharing markdown files across our team. Obsidian is the best alternative we tried, but found it doesn’t have a true “what you see is what you get” UI and it didn’t integrate well with Claude/Codex outside of community plugins. So we built OpenKnowledge. It takes shape as: 1. A MacOS app with a file navigator, the WYSIWYG editor, and link explorer. 2. Integrations with the Claude, Codex, and Cursor desktop apps. The agents can open an OpenKnowledge editor within their embedded web browsers for a side-by-side experience. 3. Built-in mcps, skills, and RAG for LLM-wiki and “AI Second Brain” scenarios + spec writing 4. An embedded terminal and CLI for TUI-first users OSS stack includes: Tiptap/prosemirror, CodeMirror, yjs (CRDT), Electron (MacOS app), Orama, remark/rehype/micromark/mdast, @pierre/trees On the architecture side, the interesting eng. challenges included: 1. A pipeline to convert ProseMirror to markdown in a bidirectional lossless way. ProseMirror uses ASTs, which are not designed to have byte-fidelity. 2. A dual-observer CRDT to keep the ProseMirror and markdown state in-sync. The CRDT + git also power a collaborative experience that shows what Agents are doing in the markdown, have undo/redo, and version history. The “Share” and cloud-sync functionality are geared for team collaboration. They feel “no-code” but leverage git/GitHub under the hood, which also means data stays fully private. In that spirit, we made OpenKnowledge open source for anybody who’s curious or who’d like to contribute. We’re actively thinking about plugins/extensibility and what’s next. If you have suggestions or feedback, would love to hear it.
27 by engomez | 5 comments on Hacker News.
Hi HN, Nick here. We’re launching OpenKnowledge ( https://ift.tt/PN4MyL5 ), a “what you see is what you get” markdown editor that has direct integrations with Claude, Codex, and Cursor. Available as MacOS app or CLI. Fully free/local and OSS ( https://ift.tt/GvhFJlw ). We built this because we wanted a “Google docs” like experience for writing and sharing markdown files across our team. Obsidian is the best alternative we tried, but found it doesn’t have a true “what you see is what you get” UI and it didn’t integrate well with Claude/Codex outside of community plugins. So we built OpenKnowledge. It takes shape as: 1. A MacOS app with a file navigator, the WYSIWYG editor, and link explorer. 2. Integrations with the Claude, Codex, and Cursor desktop apps. The agents can open an OpenKnowledge editor within their embedded web browsers for a side-by-side experience. 3. Built-in mcps, skills, and RAG for LLM-wiki and “AI Second Brain” scenarios + spec writing 4. An embedded terminal and CLI for TUI-first users OSS stack includes: Tiptap/prosemirror, CodeMirror, yjs (CRDT), Electron (MacOS app), Orama, remark/rehype/micromark/mdast, @pierre/trees On the architecture side, the interesting eng. challenges included: 1. A pipeline to convert ProseMirror to markdown in a bidirectional lossless way. ProseMirror uses ASTs, which are not designed to have byte-fidelity. 2. A dual-observer CRDT to keep the ProseMirror and markdown state in-sync. The CRDT + git also power a collaborative experience that shows what Agents are doing in the markdown, have undo/redo, and version history. The “Share” and cloud-sync functionality are geared for team collaboration. They feel “no-code” but leverage git/GitHub under the hood, which also means data stays fully private. In that spirit, we made OpenKnowledge open source for anybody who’s curious or who’d like to contribute. We’re actively thinking about plugins/extensibility and what’s next. If you have suggestions or feedback, would love to hear it.
Wednesday, June 24, 2026
New top story on Hacker News: NSA lost access to Mythos amid Anthropic dispute
NSA lost access to Mythos amid Anthropic dispute
53 by thm | 16 comments on Hacker News.
Unlocked: https://ift.tt/iGYKWLq...
53 by thm | 16 comments on Hacker News.
Unlocked: https://ift.tt/iGYKWLq...
Tuesday, June 23, 2026
Monday, June 22, 2026
New top story on Hacker News: Bain tests software takeover targets by vibecoding AI replicas
Bain tests software takeover targets by vibecoding AI replicas
9 by macleginn | 18 comments on Hacker News.
https://ift.tt/2LuFsjG
9 by macleginn | 18 comments on Hacker News.
https://ift.tt/2LuFsjG
Sunday, June 21, 2026
New top story on Hacker News: Show HN: Pulse – Dashboard for Claude Code, approve tool calls from your phone
Show HN: Pulse – Dashboard for Claude Code, approve tool calls from your phone
9 by nikitadvd | 3 comments on Hacker News.
Hi everyone, I'm a student from Flanders and I like to use Claude Code for my purposes, ideas and also just for fun and also make solutions for problems in our world!) So that's why I built "Pulse", it's an local application that you can easily install to your device and easily follow what your claude agent is doing right now in your terminal session with an ambiance design and easily give permissions for your agent. For those who wants to see directly how much tokens you spent, and how much the session costs, and approve tool calls from everywhere from your phone and everything runs locally without an account can install Pulse from GitHub: https://ift.tt/i0KWSJB
9 by nikitadvd | 3 comments on Hacker News.
Hi everyone, I'm a student from Flanders and I like to use Claude Code for my purposes, ideas and also just for fun and also make solutions for problems in our world!) So that's why I built "Pulse", it's an local application that you can easily install to your device and easily follow what your claude agent is doing right now in your terminal session with an ambiance design and easily give permissions for your agent. For those who wants to see directly how much tokens you spent, and how much the session costs, and approve tool calls from everywhere from your phone and everything runs locally without an account can install Pulse from GitHub: https://ift.tt/i0KWSJB
Saturday, June 20, 2026
Friday, June 19, 2026
Thursday, June 18, 2026
New top story on Hacker News: Agentic Resource Discovery Specification
Agentic Resource Discovery Specification
6 by damick | 0 comments on Hacker News.
https://ift.tt/ClBE8kS...
6 by damick | 0 comments on Hacker News.
https://ift.tt/ClBE8kS...
Wednesday, June 17, 2026
New top story on Hacker News: The hacker sent by Anthropic to calm the government's nerves about AI safety
The hacker sent by Anthropic to calm the government's nerves about AI safety
3 by Brajeshwar | 0 comments on Hacker News.
https://ift.tt/FrLx6Uq...
3 by Brajeshwar | 0 comments on Hacker News.
https://ift.tt/FrLx6Uq...
Tuesday, June 16, 2026
Monday, June 15, 2026
Sunday, June 14, 2026
New top story on Hacker News: Show HN: Trace – Offline Mac meeting transcripts you can flag mid-call
Show HN: Trace – Offline Mac meeting transcripts you can flag mid-call
11 by AG342 | 2 comments on Hacker News.
I'm the developer of Trace, a non-intrusive, shortcut-driven Mac app that records and transcribes your meetings on-device. I know, another meeting transcription app. Please bear with me though, I'm confident that this is at least a little novel. I primarily built Trace for myself. I'd been using MacWhisper, but there was enough fiddling before each call that I'd forget to start it and walk out of an hour-long meeting with nothing written down. So the things I cared about most were that it's quick to activate and stays out of the way. You activate Trace by pressing a global shortcut (configurable), which reveals a small bar at the bottom of your screen (there's also a keystroke and/or option to hide it entirely if you'd rather not see it at all). As I was building it I wanted to bake in a couple of workflows I'd wished for in other transcription apps. 1. Mid-meeting you can press another global shortcut to mark a "key moment" and type a note. The note shows up in the resulting transcript inline at that timestamp. I wanted to add this because I kept catching myself thinking "wait, that bit matters" in meetings and reaching to jot it down in a separate app like Obsidian, which I then needed to add context to, which took me out of the meeting. I use it all the time. If I paste the transcript into an LLM afterwards (which I find myself doing more and more these days) the important moments are flagged so it doesn't gloss over them. This is more noticeable in longer meetings with lots of topics. 2. With another keyboard shortcut you can summon a rough live recap (subtitles, basically) to quickly recap what's just been said. Trace uses standard macOS microphone and system recording APIs to capture both sides of the conversation as two separate tracks and then runs the system side through on-device diarization to identify speakers. Right now we only label them as "Speaker 1", "Speaker 2", etc but there are plans for speaker labelling in the future. You can also show a "live recap" as the call is happening to review what someone just said. All transcription models run on your machine. To be clear though, Trace doesn't do any of the summarising itself, it just produces a markdown transcript, so if you want summaries then you need to pass the output to an AI. The app is sandboxed and your audio/transcripts are never uploaded anywhere - they just exist as audio files and markdown on disk. The only network call Trace is required to make is on the first run to download the speech and speaker models (around 500MB) from Hugging Face, and after that it can be used fully offline. If enabled, a Google Calendar integration can auto-name sessions but that needs a network connection. The app is £9.99 on the macOS App Store. I've been using it every day for months now and I'm super happy with how it's improved my workflow. Feedback very welcome.
11 by AG342 | 2 comments on Hacker News.
I'm the developer of Trace, a non-intrusive, shortcut-driven Mac app that records and transcribes your meetings on-device. I know, another meeting transcription app. Please bear with me though, I'm confident that this is at least a little novel. I primarily built Trace for myself. I'd been using MacWhisper, but there was enough fiddling before each call that I'd forget to start it and walk out of an hour-long meeting with nothing written down. So the things I cared about most were that it's quick to activate and stays out of the way. You activate Trace by pressing a global shortcut (configurable), which reveals a small bar at the bottom of your screen (there's also a keystroke and/or option to hide it entirely if you'd rather not see it at all). As I was building it I wanted to bake in a couple of workflows I'd wished for in other transcription apps. 1. Mid-meeting you can press another global shortcut to mark a "key moment" and type a note. The note shows up in the resulting transcript inline at that timestamp. I wanted to add this because I kept catching myself thinking "wait, that bit matters" in meetings and reaching to jot it down in a separate app like Obsidian, which I then needed to add context to, which took me out of the meeting. I use it all the time. If I paste the transcript into an LLM afterwards (which I find myself doing more and more these days) the important moments are flagged so it doesn't gloss over them. This is more noticeable in longer meetings with lots of topics. 2. With another keyboard shortcut you can summon a rough live recap (subtitles, basically) to quickly recap what's just been said. Trace uses standard macOS microphone and system recording APIs to capture both sides of the conversation as two separate tracks and then runs the system side through on-device diarization to identify speakers. Right now we only label them as "Speaker 1", "Speaker 2", etc but there are plans for speaker labelling in the future. You can also show a "live recap" as the call is happening to review what someone just said. All transcription models run on your machine. To be clear though, Trace doesn't do any of the summarising itself, it just produces a markdown transcript, so if you want summaries then you need to pass the output to an AI. The app is sandboxed and your audio/transcripts are never uploaded anywhere - they just exist as audio files and markdown on disk. The only network call Trace is required to make is on the first run to download the speech and speaker models (around 500MB) from Hugging Face, and after that it can be used fully offline. If enabled, a Google Calendar integration can auto-name sessions but that needs a network connection. The app is £9.99 on the macOS App Store. I've been using it every day for months now and I'm super happy with how it's improved my workflow. Feedback very welcome.
Saturday, June 13, 2026
Friday, June 12, 2026
Thursday, June 11, 2026
New top story on Hacker News: Show HN: I built a Red Flag Warning zone-check tool for the East Bay in 48h
Show HN: I built a Red Flag Warning zone-check tool for the East Bay in 48h
8 by vedant28t | 0 comments on Hacker News.
Hey HN. I'm a high schooler in Fremont, CA. Tuesday morning I got a county-wide AC Alert text telling everyone in Alameda County to prepare a go-bag for an East Bay Hills Red Flag Warning that starts tonight at 11 PM. The text went to ~half a million phones. The actual NWS warning polygon only covers East Bay Hills (NWS zone CAZ515). Most people who got the text don't need a go-bag tonight. Some in the hills don't realize how close they are. So I built this tool - https://ift.tt/sLHjuqi mit licensed public github - https://ift.tt/z5Ifhli It does a few things - tells people if they are in the flagged zone, and also provides a way to check if a buddy is in flagged zone and send them a text. Everything without installing an app. I heard back from Oakland Firesafe Council director about a gap in my understanding (and the tool). To my surprise, and through feedback, I realized that you cannot assume that only the flagged area is at risk. Adjacent areas are at risk too! Fires do not follow zone boundaries! I fixed the tool. I built this in 48 hours to close that specific gap: type your address, get a yes/no on whether the NWS polygon covers it, your Genasys evacuation zone, tonight's wind + humidity at your point, a plain-English action checklist, a per-school decision view for East Bay districts, and a one-tap iMessage buddy-check template for a hill-neighbor at 10:30 PM.
8 by vedant28t | 0 comments on Hacker News.
Hey HN. I'm a high schooler in Fremont, CA. Tuesday morning I got a county-wide AC Alert text telling everyone in Alameda County to prepare a go-bag for an East Bay Hills Red Flag Warning that starts tonight at 11 PM. The text went to ~half a million phones. The actual NWS warning polygon only covers East Bay Hills (NWS zone CAZ515). Most people who got the text don't need a go-bag tonight. Some in the hills don't realize how close they are. So I built this tool - https://ift.tt/sLHjuqi mit licensed public github - https://ift.tt/z5Ifhli It does a few things - tells people if they are in the flagged zone, and also provides a way to check if a buddy is in flagged zone and send them a text. Everything without installing an app. I heard back from Oakland Firesafe Council director about a gap in my understanding (and the tool). To my surprise, and through feedback, I realized that you cannot assume that only the flagged area is at risk. Adjacent areas are at risk too! Fires do not follow zone boundaries! I fixed the tool. I built this in 48 hours to close that specific gap: type your address, get a yes/no on whether the NWS polygon covers it, your Genasys evacuation zone, tonight's wind + humidity at your point, a plain-English action checklist, a per-school decision view for East Bay districts, and a one-tap iMessage buddy-check template for a hill-neighbor at 10:30 PM.
Wednesday, June 10, 2026
Tuesday, June 9, 2026
Monday, June 8, 2026
Sunday, June 7, 2026
Saturday, June 6, 2026
Friday, June 5, 2026
Thursday, June 4, 2026
New top story on Hacker News: Show HN: Cost.dev (YC W21) – making agents cost-aware and cheaper to call
Show HN: Cost.dev (YC W21) – making agents cost-aware and cheaper to call
8 by akh | 1 comments on Hacker News.
We launched Infracost on HN five years ago ( https://ift.tt/1wrugWz ) where our CLI generated cost estimates for infra-as-code, e.g. "this Terraform PR adds $400/mo". The idea was to shift cloud costs (FinOps) left, so engineers get visibility of costs before deployment and make better decisions. Earlier this year we started seeing agent traffic in our logs and it looked like coding agents were calling our CLI. But that CLI wasn't designed with coding agents in mind. We went down a philosophical rabbit hole to see if a CLI is even needed anymore given that Claude, Copilot et al. already follow best practices. Ultimately we decided to create a new CLI from the ground up with coding agents in mind for two reasons: 1. We optimized the CLI for agent callers and cut Claude's output token usage by up to 79% and API cost by up to 67% versus a bare-Claude baseline. We wrote a blog documenting our lessons on optimizing user token usage when designing a CLI, e.g. using predicate flags so the agent doesn't compose jq | python | wc pipelines, output format that strips JSON's redundant field names. The blog is here: https://ift.tt/a3pRmOb... 2. With cloud costs, precision matters. Telling a coding agent "make this Terraform cost-optimized" can be expensive and lossy. You burn tokens loading code and policy context into every conversation. Your agent could make up a price and you wouldn't know because it's difficult to verify that across the ~10M price points that AWS, Azure and Google have. The CLI runs static analysis on the code, uses the latest prices from cloud vendors, and passes that context to the coding agent. So that's what we're launching today - Cost.dev: https://cost.dev/ . - It runs locally. Your code never leaves your machine, you get a fast feedback loop, and you're not burning API calls per character when you want to fetch prices. - The CLI does the deterministic work. Fetching price points, scanning the code, validating fixes. The coding agent does the natural-language part. You don't have to trust the LLM to remember the rules, and can verify it called the right CLI command. - It provides a consistent rule layer across every tool you use. Get cost estimates in your IDE and your coding agent with a single install. We support Claude Code, GitHub Copilot, Cursor, Windsurf, OpenAI Codex, Gemini CLI, as well as IDEs like VS Code and JetBrains Before we keep building more in that direction, I want to sanity-check with HN: is "agents writing IaC in prod" actually a thing yet, or am I betting on a future that's still a year out? I know software developers are using coding agents heavily, but are platform/infra folks doing that for prod too? Also, if you have any feedback on Cost.dev, I'd love to hear it!
8 by akh | 1 comments on Hacker News.
We launched Infracost on HN five years ago ( https://ift.tt/1wrugWz ) where our CLI generated cost estimates for infra-as-code, e.g. "this Terraform PR adds $400/mo". The idea was to shift cloud costs (FinOps) left, so engineers get visibility of costs before deployment and make better decisions. Earlier this year we started seeing agent traffic in our logs and it looked like coding agents were calling our CLI. But that CLI wasn't designed with coding agents in mind. We went down a philosophical rabbit hole to see if a CLI is even needed anymore given that Claude, Copilot et al. already follow best practices. Ultimately we decided to create a new CLI from the ground up with coding agents in mind for two reasons: 1. We optimized the CLI for agent callers and cut Claude's output token usage by up to 79% and API cost by up to 67% versus a bare-Claude baseline. We wrote a blog documenting our lessons on optimizing user token usage when designing a CLI, e.g. using predicate flags so the agent doesn't compose jq | python | wc pipelines, output format that strips JSON's redundant field names. The blog is here: https://ift.tt/a3pRmOb... 2. With cloud costs, precision matters. Telling a coding agent "make this Terraform cost-optimized" can be expensive and lossy. You burn tokens loading code and policy context into every conversation. Your agent could make up a price and you wouldn't know because it's difficult to verify that across the ~10M price points that AWS, Azure and Google have. The CLI runs static analysis on the code, uses the latest prices from cloud vendors, and passes that context to the coding agent. So that's what we're launching today - Cost.dev: https://cost.dev/ . - It runs locally. Your code never leaves your machine, you get a fast feedback loop, and you're not burning API calls per character when you want to fetch prices. - The CLI does the deterministic work. Fetching price points, scanning the code, validating fixes. The coding agent does the natural-language part. You don't have to trust the LLM to remember the rules, and can verify it called the right CLI command. - It provides a consistent rule layer across every tool you use. Get cost estimates in your IDE and your coding agent with a single install. We support Claude Code, GitHub Copilot, Cursor, Windsurf, OpenAI Codex, Gemini CLI, as well as IDEs like VS Code and JetBrains Before we keep building more in that direction, I want to sanity-check with HN: is "agents writing IaC in prod" actually a thing yet, or am I betting on a future that's still a year out? I know software developers are using coding agents heavily, but are platform/infra folks doing that for prod too? Also, if you have any feedback on Cost.dev, I'd love to hear it!
Wednesday, June 3, 2026
New top story on Hacker News: Launch HN: Hyper (YC P26) – Company brain to power agentic development
Launch HN: Hyper (YC P26) – Company brain to power agentic development
16 by shalinshah | 6 comments on Hacker News.
Hey HN, we’re Shalin & Kanyes, best friends who've been hacking together for 10+yrs, and now founders of Hyper ( https://heyhyper.ai/ ). Hyper is a shared “company brain” that plugs into information flowing inside a company to make AI agents and automations better and ultimately save people time. Models have gotten good enough that they can (mostly) take on long-horizon, complex tasks. We believe the bottleneck now is that these smart-enough models often lack information about your company, which is scattered in people's heads, Slack threads, stale docs, and in back-and-forth convos with AI. MCP is useful for getting some info in front of an agent, but there are problems: (1) Once the session dies, so does the insight, so instead of copy-pasting a whole doc each time you're telling the agent to dig through Drive each time - not much of a win; (2) Even when MCP works, what it gathers isn't comprehensive, because people decide things on a whiteboard, brainstorm out loud, post a little in Slack, and scribble the rest in a doc, which leaves the agent working from partial information; (3) And even if it had everything, it doesn't do the meta-reasoning required to do a great job. If you paste in a Notion doc and it won't learn your design taste or your writing style unless you tell it to, and it won't know why a decision was made or when. As undergrads 5 years ago, we were into the tools-for-thought wave and became power users of Notion, Obsidian, Roam, Anki, real believers in building a second brain. After GPT-3.5 came out we started to realize how much more powerful that second brain could be if an AI could actually read it, because suddenly it would know our backstory, our taste, our preferences, and unlock genuinely new capabilities. That’s why we’re building Hyper. We know it’s not for everybody! But for people who do want to be on the cutting edge, this is a force multiplier that makes agents faster and better. It increases the number of tasks they can do, and how effectively they do them. Hyper works by ingesting everything you give it access to, Docs, Slack, Email, Calendar, Granola, and synthesizes it into a knowledge graph of facts and their relationships with embeddings for semantic search. The memory system we’ve built is hybrid, with two modalities. Episodes are the raw source items kept as the source of truth. Facts are the meaning pulled out of each episode, stored as subject-predicate-object records with a plain summary and timestamps for when the fact was introduced and when it was invalidated (subject=person, predicate=works_at, object=company). Facts form a graph with typed edges between them: X is in tension with Y, A is derived from B, J supersedes K. Every time a new fact comes in we update the facts in its neighborhood, so the graph stays current, and that's how we handle stale information. When "we'll ship Friday" is later contradicted by "we're shipping Monday," the new fact supersedes the old one instead of both looking equally true, and we never auto-discard the superseded version, so you can still ask how you landed on Monday. Every fact carries provenance back to its source and access-control tags for who is allowed to see it. At retrieval we query-expand, then fuse semantic search over embeddings with Postgres full-text search using reciprocal rank fusion, and we only ever evaluate a query against the facts and episodes that person has access to, which means two people on the same team can ask the same question and get different answers. We keep information fresh with webhooks where they exist and polling where they don't, hashing contents to catch changes for sources that don’t handle native dedupe. Agents read and write through two paths: lifecycle hooks in tools like Claude Code, Cowork, Codex, and Cursor, where we inject relevant context on every prompt and pull interesting facts out of every response, and plain MCP tool calls for everything that doesn't expose hooks. We love it! and so do our early users: one CEO uses Hyper to draft emails in his voice with full company context. What took hours/week now takes minutes and gets sharper each time Hyper learns more how he thinks and how his company is changing. Another YC founder one-shotted a launch video script because Hyper already knew their product, voice, positioning accumulated over months. We have a 3-day free trial, explained more on our pricing page ( https://ift.tt/z23S1fc ) and there are more details in our FAQ ( https://heyhyper.ai/faq ), including things like privacy, compliance, and how we’re different from other “memory” companies.. Give it a spin! break it! and tell us where it falls short: https://heyhyper.ai/ . We'd love to build you a 10-star experience :) Comments welcome!
16 by shalinshah | 6 comments on Hacker News.
Hey HN, we’re Shalin & Kanyes, best friends who've been hacking together for 10+yrs, and now founders of Hyper ( https://heyhyper.ai/ ). Hyper is a shared “company brain” that plugs into information flowing inside a company to make AI agents and automations better and ultimately save people time. Models have gotten good enough that they can (mostly) take on long-horizon, complex tasks. We believe the bottleneck now is that these smart-enough models often lack information about your company, which is scattered in people's heads, Slack threads, stale docs, and in back-and-forth convos with AI. MCP is useful for getting some info in front of an agent, but there are problems: (1) Once the session dies, so does the insight, so instead of copy-pasting a whole doc each time you're telling the agent to dig through Drive each time - not much of a win; (2) Even when MCP works, what it gathers isn't comprehensive, because people decide things on a whiteboard, brainstorm out loud, post a little in Slack, and scribble the rest in a doc, which leaves the agent working from partial information; (3) And even if it had everything, it doesn't do the meta-reasoning required to do a great job. If you paste in a Notion doc and it won't learn your design taste or your writing style unless you tell it to, and it won't know why a decision was made or when. As undergrads 5 years ago, we were into the tools-for-thought wave and became power users of Notion, Obsidian, Roam, Anki, real believers in building a second brain. After GPT-3.5 came out we started to realize how much more powerful that second brain could be if an AI could actually read it, because suddenly it would know our backstory, our taste, our preferences, and unlock genuinely new capabilities. That’s why we’re building Hyper. We know it’s not for everybody! But for people who do want to be on the cutting edge, this is a force multiplier that makes agents faster and better. It increases the number of tasks they can do, and how effectively they do them. Hyper works by ingesting everything you give it access to, Docs, Slack, Email, Calendar, Granola, and synthesizes it into a knowledge graph of facts and their relationships with embeddings for semantic search. The memory system we’ve built is hybrid, with two modalities. Episodes are the raw source items kept as the source of truth. Facts are the meaning pulled out of each episode, stored as subject-predicate-object records with a plain summary and timestamps for when the fact was introduced and when it was invalidated (subject=person, predicate=works_at, object=company). Facts form a graph with typed edges between them: X is in tension with Y, A is derived from B, J supersedes K. Every time a new fact comes in we update the facts in its neighborhood, so the graph stays current, and that's how we handle stale information. When "we'll ship Friday" is later contradicted by "we're shipping Monday," the new fact supersedes the old one instead of both looking equally true, and we never auto-discard the superseded version, so you can still ask how you landed on Monday. Every fact carries provenance back to its source and access-control tags for who is allowed to see it. At retrieval we query-expand, then fuse semantic search over embeddings with Postgres full-text search using reciprocal rank fusion, and we only ever evaluate a query against the facts and episodes that person has access to, which means two people on the same team can ask the same question and get different answers. We keep information fresh with webhooks where they exist and polling where they don't, hashing contents to catch changes for sources that don’t handle native dedupe. Agents read and write through two paths: lifecycle hooks in tools like Claude Code, Cowork, Codex, and Cursor, where we inject relevant context on every prompt and pull interesting facts out of every response, and plain MCP tool calls for everything that doesn't expose hooks. We love it! and so do our early users: one CEO uses Hyper to draft emails in his voice with full company context. What took hours/week now takes minutes and gets sharper each time Hyper learns more how he thinks and how his company is changing. Another YC founder one-shotted a launch video script because Hyper already knew their product, voice, positioning accumulated over months. We have a 3-day free trial, explained more on our pricing page ( https://ift.tt/z23S1fc ) and there are more details in our FAQ ( https://heyhyper.ai/faq ), including things like privacy, compliance, and how we’re different from other “memory” companies.. Give it a spin! break it! and tell us where it falls short: https://heyhyper.ai/ . We'd love to build you a 10-star experience :) Comments welcome!
Tuesday, June 2, 2026
New top story on Hacker News: Show HN: RePlaya – self-hosted browser session replay with live tailing
Show HN: RePlaya – self-hosted browser session replay with live tailing
4 by shikhar | 0 comments on Hacker News.
Hi HN, I'm one of the founders of s2.dev. RePlaya ( https://ift.tt/6yGE3X5 ) is a self-hosted browser session replay tool using rrweb ( https://ift.tt/2oas0K5 ). It occurred to me that a durable stream per session would be a much neater architectural foundation for much of what you'd want from such a tool. As a unique feature, it also made live tailing straightforward because the player can read from the same stream the recorder is appending to. The alternative architecture is likely an ingest firehose which is then indexed, with associated complexity and latency. You'd have to string together multiple data systems like a message queue, a metadata database, and blob storage and/or an OLAP database. Here the only dependency is S2, which has an open source version you can self-host called s2-lite ( https://ift.tt/LOkswu2 ). How it works: - one S2 stream per browser session - large rrweb events (like a full snapshot) get framed across multiple binary S2 records and reassembled on read - active sessions are tailed with an S2 read session, and bridged to the browser over SSE - session listing relies on stream names encoding reverse timestamps, as S2 returns a lexicographic order listing - relying on fencing tokens so a stopped session can't be written to again by a late recorder - retention and GC are handled via S2 stream config, so no background job needed Curious to hear from folks on the tool or the stream-per-session model!
4 by shikhar | 0 comments on Hacker News.
Hi HN, I'm one of the founders of s2.dev. RePlaya ( https://ift.tt/6yGE3X5 ) is a self-hosted browser session replay tool using rrweb ( https://ift.tt/2oas0K5 ). It occurred to me that a durable stream per session would be a much neater architectural foundation for much of what you'd want from such a tool. As a unique feature, it also made live tailing straightforward because the player can read from the same stream the recorder is appending to. The alternative architecture is likely an ingest firehose which is then indexed, with associated complexity and latency. You'd have to string together multiple data systems like a message queue, a metadata database, and blob storage and/or an OLAP database. Here the only dependency is S2, which has an open source version you can self-host called s2-lite ( https://ift.tt/LOkswu2 ). How it works: - one S2 stream per browser session - large rrweb events (like a full snapshot) get framed across multiple binary S2 records and reassembled on read - active sessions are tailed with an S2 read session, and bridged to the browser over SSE - session listing relies on stream names encoding reverse timestamps, as S2 returns a lexicographic order listing - relying on fencing tokens so a stopped session can't be written to again by a late recorder - retention and GC are handled via S2 stream config, so no background job needed Curious to hear from folks on the tool or the stream-per-session model!
New top story on Hacker News: Trump signs downsized AI order after weeks of reversals
Trump signs downsized AI order after weeks of reversals
33 by _alternator_ | 16 comments on Hacker News.
https://ift.tt/aoGUtEY... https://ift.tt/fR0xvYL...
33 by _alternator_ | 16 comments on Hacker News.
https://ift.tt/aoGUtEY... https://ift.tt/fR0xvYL...
Monday, June 1, 2026
New top story on Hacker News: Superintelligence: The Idea That Eats Smart People (2016)
Superintelligence: The Idea That Eats Smart People (2016)
20 by thoughtpeddler | 3 comments on Hacker News.
20 by thoughtpeddler | 3 comments on Hacker News.
Sunday, May 31, 2026
Saturday, May 30, 2026
Friday, May 29, 2026
New top story on Hacker News: ATLAS: Autoformalized Textbook Library At Scale
ATLAS: Autoformalized Textbook Library At Scale
15 by vrm | 1 comments on Hacker News.
https://twitter.com/arnal_charles/status/2060009395107377282 , https://ift.tt/LfWZC2O Paper: Formalizing Mathematics at Scale - https://ift.tt/NLerXMg
15 by vrm | 1 comments on Hacker News.
https://twitter.com/arnal_charles/status/2060009395107377282 , https://ift.tt/LfWZC2O Paper: Formalizing Mathematics at Scale - https://ift.tt/NLerXMg
Thursday, May 28, 2026
New top story on Hacker News: Show HN: Ktx – Open-source executable context layer for data agents
Show HN: Ktx – Open-source executable context layer for data agents
16 by lucamrtl | 3 comments on Hacker News.
Hi HN, we’re open-sourcing ktx. It’s an executable context layer that makes agents reliable on your data stack. We built it after going through the experience of building production-grade data agents for dozens of companies. If you’ve also tried building them, or simply tried using Claude Code or Codex on your data warehouse, you’ll know that accuracy is the #1 issue. Agents are great at generating valid SQL, but it’s not always correct SQL. To cite a few examples of “agents gone wrong”: - Stale column + hidden business rule: when preparing a board report, a finance analyst asks Claude Code for “ARR by customer segment”, it derives ARR from multiple tables (subscriptions, plans, accounts), then groups by accounts.industry. But CC doesn’t know that this industry column was deprecated a few months prior, or that past board reports excluded paused subscriptions from the ARR calculation - Join fanout: a data analyst at a retailer uses their company’s internal agent to prep a product revenue deck for a QBR. The agent joins orders to order_items, then sums orders.total_amount_cents grouped by order_items.product_id. The SQL runs fine, but each order’s revenue is repeated once per line item, which most people will miss if most orders only have 1 item - Missing attribution logic: a marketing analyst asks Codex “Which campaigns drove the most revenue?” Codex joins marketing_touches to users to orders and groups by utm_campaign. But since each order can have multiple touches before purchase, the same order can be credited to first touch, last touch, every touch, or every campaign the user clicked before buying. If the agent chooses the method that doesn’t match the team’s attribution logic, they’ll make suboptimal decisions To solve this at first we gave the agent more context through skills + a wiki-style knowledge base. That gives it some useful extra context but still relies on it writing the SQL without incorrect assumptions. The next solution we explored was implementing a classic semantic layer. That solves the executable part, but they’re such a pain to build and maintain since they were made for legacy BI tools. Plus as a standalone tool, they lack all the useful context from unstructured data sources like internal docs. So we built ktx and split it into 2 parts: 1. Business context goes in Markdown wiki pages that are auto-ingested and auto-populated 2. Queryable definitions go into YAML files that define tables, row grain, joins, measures, dimensions, filters, and filter groups That way, when an agent needs a metric, it asks ktx for a measure, dimensions, filters, and filter groups instead of writing the whole query itself. ktx’s planner chooses the join path, uses grain and relationship metadata, catches issues like join fanout and chasm joins, and compiles the warehouse SQL, while utilizing the extra unstructured knowledge it has access to. ktx is Apache 2.0. It can ingest from most warehouses (BigQuery, Snowflake, Postgres & others), modeling tools (dbt, MetricFlow, LookML), BI tools (Looker, Metabase), doc tools like Notion, and corrections from user interactions. Install manually: npm install -g @kaelio/ktx ktx setup Or give this prompt to your agent: Run npx skills add Kaelio/ktx --skill ktx and use ktx skill to install and configure ktx We’d especially like feedback from people who’ve tried using Claude Code, Codex, or building custom agents on analytics warehouses. Where did they fail? And what did you try to make the answers more reliable?
16 by lucamrtl | 3 comments on Hacker News.
Hi HN, we’re open-sourcing ktx. It’s an executable context layer that makes agents reliable on your data stack. We built it after going through the experience of building production-grade data agents for dozens of companies. If you’ve also tried building them, or simply tried using Claude Code or Codex on your data warehouse, you’ll know that accuracy is the #1 issue. Agents are great at generating valid SQL, but it’s not always correct SQL. To cite a few examples of “agents gone wrong”: - Stale column + hidden business rule: when preparing a board report, a finance analyst asks Claude Code for “ARR by customer segment”, it derives ARR from multiple tables (subscriptions, plans, accounts), then groups by accounts.industry. But CC doesn’t know that this industry column was deprecated a few months prior, or that past board reports excluded paused subscriptions from the ARR calculation - Join fanout: a data analyst at a retailer uses their company’s internal agent to prep a product revenue deck for a QBR. The agent joins orders to order_items, then sums orders.total_amount_cents grouped by order_items.product_id. The SQL runs fine, but each order’s revenue is repeated once per line item, which most people will miss if most orders only have 1 item - Missing attribution logic: a marketing analyst asks Codex “Which campaigns drove the most revenue?” Codex joins marketing_touches to users to orders and groups by utm_campaign. But since each order can have multiple touches before purchase, the same order can be credited to first touch, last touch, every touch, or every campaign the user clicked before buying. If the agent chooses the method that doesn’t match the team’s attribution logic, they’ll make suboptimal decisions To solve this at first we gave the agent more context through skills + a wiki-style knowledge base. That gives it some useful extra context but still relies on it writing the SQL without incorrect assumptions. The next solution we explored was implementing a classic semantic layer. That solves the executable part, but they’re such a pain to build and maintain since they were made for legacy BI tools. Plus as a standalone tool, they lack all the useful context from unstructured data sources like internal docs. So we built ktx and split it into 2 parts: 1. Business context goes in Markdown wiki pages that are auto-ingested and auto-populated 2. Queryable definitions go into YAML files that define tables, row grain, joins, measures, dimensions, filters, and filter groups That way, when an agent needs a metric, it asks ktx for a measure, dimensions, filters, and filter groups instead of writing the whole query itself. ktx’s planner chooses the join path, uses grain and relationship metadata, catches issues like join fanout and chasm joins, and compiles the warehouse SQL, while utilizing the extra unstructured knowledge it has access to. ktx is Apache 2.0. It can ingest from most warehouses (BigQuery, Snowflake, Postgres & others), modeling tools (dbt, MetricFlow, LookML), BI tools (Looker, Metabase), doc tools like Notion, and corrections from user interactions. Install manually: npm install -g @kaelio/ktx ktx setup Or give this prompt to your agent: Run npx skills add Kaelio/ktx --skill ktx and use ktx skill to install and configure ktx We’d especially like feedback from people who’ve tried using Claude Code, Codex, or building custom agents on analytics warehouses. Where did they fail? And what did you try to make the answers more reliable?
Wednesday, May 27, 2026
Tuesday, May 26, 2026
New top story on Hacker News: Dropbox CEO Drew Houston to step down
Dropbox CEO Drew Houston to step down
83 by aghuang | 46 comments on Hacker News.
https://ift.tt/18B6hWI...
83 by aghuang | 46 comments on Hacker News.
https://ift.tt/18B6hWI...
Monday, May 25, 2026
Sunday, May 24, 2026
Saturday, May 23, 2026
Friday, May 22, 2026
Thursday, May 21, 2026
Wednesday, May 20, 2026
New top story on Hacker News: Show HN: Hocuspocus 4 – self-hosted Yjs collaboration backend
Show HN: Hocuspocus 4 – self-hosted Yjs collaboration backend
11 by philipisik | 3 comments on Hacker News.
Hi HN! I'm Philip, one of the founders of Tiptap. Next to our open-source rich text editor framework, we started developing Hocuspocus about five years ago and open-sourced it too, to solve one of our biggest challenges back then: real-time collaboration in web editors. We found Yjs by Kevin Jahns, a CRDT library that handles concurrent edits without conflicts. Basically, Yjs merges changes from users without conflicts and in real-time. Hocuspocus is the WebSocket server built on top of Yjs. It handles real-time sync, presence/awareness, persistence, and Redis-based scaling. While we use Hocuspocus at Tiptap as the collaboration backend for our cloud services, it also works with any Yjs client (Slate, Quill, Monaco, ProseMirror, or your own setup), and Yjs documents aren't limited to text at all. You can sync any structured data through them, and in the meantime we see projects that rely on Hocuspocus without using the Tiptap editor. We released Hocuspocus v4 under the MIT license a few weeks ago, and the biggest change is that it's no longer tied to Node. The previous versions depended on the ws package, which meant you couldn't run Hocuspocus on Bun, Deno, or Cloudflare Workers. We moved to crossws, a universal websocket adapter, so the same server now runs on Node, Bun, Deno, Cloudflare Workers, and Node with uWebSockets. That also lets you run collaboration at the edge. The other changes are smaller but are important if you're using Hocuspocus in production: 1. Every core class and hook payload takes a generic Context type now, so the auth/session shape you build in onAuthenticate flows through every other hook with full type safety (defaults to any so existing code doesn't break). 2. Document updates are now processed sequentially per connection through an internal queue, which fixes a correctness bug where async hooks could cause CRDT updates to apply out of order under load. 3. Transaction origins are structured objects now with a source field instead of raw values and there's an isTransactionOrigin() helper for narrowing. 4. Hook payloads use web-standard Request and Headers instead of Node's IncomingMessage. 5. The wire protocol is backward compatible in both directions, so you can roll out servers and providers independently. If you want to test Hocuspocus: npm install @hocuspocus/server @hocuspocus/provider Docs at: https://ift.tt/ATQLu6l Source at: https://ift.tt/R0XxyU9 Because running real-time collaboration on Workers or Durable Objects is new in v4, that's the use case we'd most like to hear your questions and feedback on.
11 by philipisik | 3 comments on Hacker News.
Hi HN! I'm Philip, one of the founders of Tiptap. Next to our open-source rich text editor framework, we started developing Hocuspocus about five years ago and open-sourced it too, to solve one of our biggest challenges back then: real-time collaboration in web editors. We found Yjs by Kevin Jahns, a CRDT library that handles concurrent edits without conflicts. Basically, Yjs merges changes from users without conflicts and in real-time. Hocuspocus is the WebSocket server built on top of Yjs. It handles real-time sync, presence/awareness, persistence, and Redis-based scaling. While we use Hocuspocus at Tiptap as the collaboration backend for our cloud services, it also works with any Yjs client (Slate, Quill, Monaco, ProseMirror, or your own setup), and Yjs documents aren't limited to text at all. You can sync any structured data through them, and in the meantime we see projects that rely on Hocuspocus without using the Tiptap editor. We released Hocuspocus v4 under the MIT license a few weeks ago, and the biggest change is that it's no longer tied to Node. The previous versions depended on the ws package, which meant you couldn't run Hocuspocus on Bun, Deno, or Cloudflare Workers. We moved to crossws, a universal websocket adapter, so the same server now runs on Node, Bun, Deno, Cloudflare Workers, and Node with uWebSockets. That also lets you run collaboration at the edge. The other changes are smaller but are important if you're using Hocuspocus in production: 1. Every core class and hook payload takes a generic Context type now, so the auth/session shape you build in onAuthenticate flows through every other hook with full type safety (defaults to any so existing code doesn't break). 2. Document updates are now processed sequentially per connection through an internal queue, which fixes a correctness bug where async hooks could cause CRDT updates to apply out of order under load. 3. Transaction origins are structured objects now with a source field instead of raw values and there's an isTransactionOrigin() helper for narrowing. 4. Hook payloads use web-standard Request and Headers instead of Node's IncomingMessage. 5. The wire protocol is backward compatible in both directions, so you can roll out servers and providers independently. If you want to test Hocuspocus: npm install @hocuspocus/server @hocuspocus/provider Docs at: https://ift.tt/ATQLu6l Source at: https://ift.tt/R0XxyU9 Because running real-time collaboration on Workers or Durable Objects is new in v4, that's the use case we'd most like to hear your questions and feedback on.
Tuesday, May 19, 2026
Monday, May 18, 2026
Sunday, May 17, 2026
Saturday, May 16, 2026
Friday, May 15, 2026
Thursday, May 14, 2026
Wednesday, May 13, 2026
Tuesday, May 12, 2026
New top story on Hacker News: Show HN: Statewright – Visual state machines that make AI agents reliable
Show HN: Statewright – Visual state machines that make AI agents reliable
8 by azurewraith | 0 comments on Hacker News.
Agentic problem solving in its current state is very brittle. I fell in love with it, but it creates as many problems as it solves. I'm Ben Cochran, I spent 20+ years in the trenches with full-stack Engineering, DevOps, high performance computing & ML with stints at NVIDIA, AMD and various other organizations most recently as a Distinguished Engineer. For agents to work reliably you either need massive parameter counts or massive context windows to keep the solution spaces workable. Most people are brute forcing reliability with bigger models and longer prompts. What if I made the problem smaller instead of making the model bigger? I took a different approach by using smaller models: models in the 13-20B parameter range and set them to task solving real SWE-bench problems. I constrained the tool and solution spaces using formal state machines. Each state in the machine defines which tools the model can access, how many iterations it gets and what transitions are valid. A planning state gets read-only tools. An implementation state gets edit tools (scoped to prevent mega edits) and write friendly bash tools. The testing state gets bash but only for testing commands. The model cannot physically skip steps or use the wrong tool at the wrong time. It is enforced via protocol, not via prompts. The results were more promising than I would have expected. Across multiple model families irrespective of age (qwen-coder, gpt-oss, gemma4) and the improvements were consistent above the 13B parameter inflection point. Below that, models can navigate the state machine but can't retain enough context to produce accurate edits. More on the research bit: https://ift.tt/2foS1dF Surprisingly this yielded improvements in frontier models as well. Haiku and Sonnet start to punch above their weight and Opus solves more reliably with fewer tokens and death spirals. Fine tuning did not yield these kinds of functional improvements for me. The takeaway it seems is that context window utilization matters more than raw context size - a tightly scoped working context at each step outperforms a model given carte blanche over everything. Constraining LLMs which are non-idempotent by using deterministic code is a pattern that nobody is currently talking about. So, I built Statewright. Its core is a Rust engine that evaluates state machine definitions: states, transitions, guards and tool restrictions. Its orchestration doesn't use an LLM, just enforces the state machine. On top of that is a plugin layer that integrates with Claude Code (and soon Codex, Cursor and others) via MCP. When you activate a workflow, hooks enforce the guardrails per state automatically. The model sees 5 tools available instead of dozens, gets clear instructions for the current phase and transitions when conditions are met. Importantly it tells the model when it's attempting to do something that isn't in scope, incorrect or when it needs to try something else after getting stuck. You can use your agent via MCP to build a state machine for you to solve a problem in your current context. The visual editor at statewright.ai lets you tweak these workflows in a graph view... You can clearly see the failure paths, the retry loops and the approval gates. State machines aren't DAGs; they loop and retry, which is what agentic work actually needs. Statewright is currently live with a free tier, try it out in Claude Code by running the following: /plugin marketplace add statewright/statewright /plugin install statewright /reload-plugins Then "start the bugfix workflow" or /statewright start bugfix. You'll need to paste your API key when prompted. The latest versions of Claude may complain -- paste the API key again and say you really mean it, Claude is just being cautious here. Feedback is welcome on the workflow editor, the plugin experience, and tell me what workflows you'd want to build first. Agents are suggestions, states are laws.
8 by azurewraith | 0 comments on Hacker News.
Agentic problem solving in its current state is very brittle. I fell in love with it, but it creates as many problems as it solves. I'm Ben Cochran, I spent 20+ years in the trenches with full-stack Engineering, DevOps, high performance computing & ML with stints at NVIDIA, AMD and various other organizations most recently as a Distinguished Engineer. For agents to work reliably you either need massive parameter counts or massive context windows to keep the solution spaces workable. Most people are brute forcing reliability with bigger models and longer prompts. What if I made the problem smaller instead of making the model bigger? I took a different approach by using smaller models: models in the 13-20B parameter range and set them to task solving real SWE-bench problems. I constrained the tool and solution spaces using formal state machines. Each state in the machine defines which tools the model can access, how many iterations it gets and what transitions are valid. A planning state gets read-only tools. An implementation state gets edit tools (scoped to prevent mega edits) and write friendly bash tools. The testing state gets bash but only for testing commands. The model cannot physically skip steps or use the wrong tool at the wrong time. It is enforced via protocol, not via prompts. The results were more promising than I would have expected. Across multiple model families irrespective of age (qwen-coder, gpt-oss, gemma4) and the improvements were consistent above the 13B parameter inflection point. Below that, models can navigate the state machine but can't retain enough context to produce accurate edits. More on the research bit: https://ift.tt/2foS1dF Surprisingly this yielded improvements in frontier models as well. Haiku and Sonnet start to punch above their weight and Opus solves more reliably with fewer tokens and death spirals. Fine tuning did not yield these kinds of functional improvements for me. The takeaway it seems is that context window utilization matters more than raw context size - a tightly scoped working context at each step outperforms a model given carte blanche over everything. Constraining LLMs which are non-idempotent by using deterministic code is a pattern that nobody is currently talking about. So, I built Statewright. Its core is a Rust engine that evaluates state machine definitions: states, transitions, guards and tool restrictions. Its orchestration doesn't use an LLM, just enforces the state machine. On top of that is a plugin layer that integrates with Claude Code (and soon Codex, Cursor and others) via MCP. When you activate a workflow, hooks enforce the guardrails per state automatically. The model sees 5 tools available instead of dozens, gets clear instructions for the current phase and transitions when conditions are met. Importantly it tells the model when it's attempting to do something that isn't in scope, incorrect or when it needs to try something else after getting stuck. You can use your agent via MCP to build a state machine for you to solve a problem in your current context. The visual editor at statewright.ai lets you tweak these workflows in a graph view... You can clearly see the failure paths, the retry loops and the approval gates. State machines aren't DAGs; they loop and retry, which is what agentic work actually needs. Statewright is currently live with a free tier, try it out in Claude Code by running the following: /plugin marketplace add statewright/statewright /plugin install statewright /reload-plugins Then "start the bugfix workflow" or /statewright start bugfix. You'll need to paste your API key when prompted. The latest versions of Claude may complain -- paste the API key again and say you really mean it, Claude is just being cautious here. Feedback is welcome on the workflow editor, the plugin experience, and tell me what workflows you'd want to build first. Agents are suggestions, states are laws.
Monday, May 11, 2026
Sunday, May 10, 2026
Saturday, May 9, 2026
Friday, May 8, 2026
Thursday, May 7, 2026
New top story on Hacker News: Show HN: Stage CLI – an easier way of reading your AI generated changes locally
Show HN: Stage CLI – an easier way of reading your AI generated changes locally
8 by cpan22 | 2 comments on Hacker News.
Hey HN! We're Charles and Dean. A few weeks ago we posted about Stage, a code review tool that guides you through reading a PR step by step - https://ift.tt/Kf1pD4I . We got a lot of great feedback but also heard from many people that they wanted to have the chapters experience even before opening a PR… so we built the Stage CLI as the local, open-source version that anyone can try. Here’s a quick demo video: https://ift.tt/qWrjyoQ It works with any coding agent of your choice. The skill instructs the agent to read your current branch’s changes, break them down into separate logical chapters, and open them in a local browser. We’ve found that reading changes this way is a lot easier for us than reading them in an IDE or other similar CLI tools, which present diffs to you in repository tree order. You can see a few examples of what it feels like here: https://ift.tt/pY0zi1k . Try it out and let us know what you think! Would love to hear any feedback :)
8 by cpan22 | 2 comments on Hacker News.
Hey HN! We're Charles and Dean. A few weeks ago we posted about Stage, a code review tool that guides you through reading a PR step by step - https://ift.tt/Kf1pD4I . We got a lot of great feedback but also heard from many people that they wanted to have the chapters experience even before opening a PR… so we built the Stage CLI as the local, open-source version that anyone can try. Here’s a quick demo video: https://ift.tt/qWrjyoQ It works with any coding agent of your choice. The skill instructs the agent to read your current branch’s changes, break them down into separate logical chapters, and open them in a local browser. We’ve found that reading changes this way is a lot easier for us than reading them in an IDE or other similar CLI tools, which present diffs to you in repository tree order. You can see a few examples of what it feels like here: https://ift.tt/pY0zi1k . Try it out and let us know what you think! Would love to hear any feedback :)
Wednesday, May 6, 2026
Tuesday, May 5, 2026
Monday, May 4, 2026
Sunday, May 3, 2026
Saturday, May 2, 2026
Friday, May 1, 2026
Thursday, April 30, 2026
Wednesday, April 29, 2026
Tuesday, April 28, 2026
Monday, April 27, 2026
New top story on Hacker News: China blocks Meta's acquisition of AI startup Manus
China blocks Meta's acquisition of AI startup Manus
33 by yakkomajuri | 7 comments on Hacker News.
https://ift.tt/iBFlkr2... https://ift.tt/IApTbhR
33 by yakkomajuri | 7 comments on Hacker News.
https://ift.tt/iBFlkr2... https://ift.tt/IApTbhR