| View in browser: The Diff No. 009 |
| The Diff No. 009 Monday, 12 October 2026 · AI dev news, only what is verified |
 |
He reran the Rust-vs-Elixir benchmark. Both lost. Piotr Sarnacki read the benchmark fight over DHH’s agent-written Campfire rewrites, saw that it only measured throughput, and reran it at a constant request rate. Rust delivered about 14% of events, Elixir about 60%, and Elixir timed out on about 23% of posts. His point: you still have to think. Today’s other threads: a relay that lets coding agents use your real signed-in Chrome, a local search index that claims 43% fewer tool calls, and a free dashboard that turns your agent sessions into a stand-up call. If your prompt does not decide it, the agent flips a coin. |
IN THIS ISSUE | 01 ⚖️ Same prompt, three rewrites Piotr Sarnacki reran the benchmark that crowned a language. | | 02 🌐 Your real Chrome, in a background tab An agent with your logins. Read the audit trail first. |  | 03 🔎 One search instead of forty reads Pensieve gives your coding agent an index, not a grep loop. |  | 04 📞 Your agents, on a conference call Agent Standup reads the session files you already have. |
|
🔥 VIRAL The loud ones on Hacker News. Numbers are a morning snapshot. | 01 HN: 332 points · post Oct 11 ⚖️ Same prompt, three rewritesPiotr Sarnacki reran the benchmark that crowned a language. SCENE Piotr Sarnacki says DHH had an agent rewrite Campfire Once from Ruby on Rails into Rust, then into Elixir and Go, without reading the code himself. Sarnacki’s post says the rewrites made different choices. The Rust one drops the CSRF token and Redis, the Elixir one stays closer to the original. People then argued which language was better from a closed-loop throughput test. THE THING Sarnacki reran the test at a constant arrival rate, which is closer to many users acting at once. At 100 posts a second he reports the Rust version delivered about 14% of events and the Elixir one about 60%, but Elixir timed out on about 23% of HTTP posts, and worst-case delivery lag reached about 180 seconds in one case. He also notes that some Rust database calls block the async runtime. These are his numbers from his own reruns. We did not run them. ✂️ STEAL IT When you ask an agent for a rewrite, write down the answers to the questions that have no single right answer: backwards compatibility, memory, loss tolerance, latency or throughput. Test with the load shape your users produce, not the one that is easy to script. Try today VERDICT Cheap to apply: list the open decisions before the prompt. It does not tell you which language wins. His runs are one author’s, on patched versions of both rewrites. Sarnacki’s post · HN discussion Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes. More like thisLess like thisDig deeperAlready knew itTest it for me |
02 253 stars · MIT · repo created Oct 10 · card: GitHub 🌐 Your real Chrome, in a background tabAn agent with your logins. Read the audit trail first. SCENE Alexey Indeev published Chrome Relay on Saturday. The README says it lets Claude Code, Codex, Cursor or Paseo drive your everyday Chrome, with your real logins, from the command line. Tabs open inactive in a collapsed Agents group, so nothing steals focus. THE THING It is a Go relay plus a Chrome extension. The agent connects through the agent-browser CLI to a local endpoint at 127.0.0.1:9333. The README lists two checks: a secret generated at setup, and the connecting process must look like Claude Code, Codex, Cursor or Paseo. Each agent only sees tabs it opened. Every click, fill and screenshot lands in a local audit log, and typed text is logged as a character count only. The README also says plainly that this is not a boundary against malware already running as you. macOS only, built and signed on your own Mac with Go. All README claims. We did not install it. ✂️ STEAL IT Even if you never install it, copy the audit idea: give an agent browser its own tab group and keep a log of what it touched that you can read afterward. Watch VERDICT The idea is useful and the README is candid about limits. But handing an agent your signed-in sessions is a large trust decision, and the project is two days old with no outside review we could find. Repo + README · agent-browser (dependency) Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes. More like thisLess like thisDig deeperAlready knew it |
🔭 UNDER THE RADAR Quiet tools and field reports worth your time. | 03 10 stars · MIT · repo created Oct 8 · HN: 37 points · card: GitHub 🔎 One search instead of forty readsPensieve gives your coding agent an index, not a grep loop. SCENE Ted Haley says on HN that when Claude Code or Codex needs code, it greps, reads a few hits, and greps again, and you pay tokens for every round. He built Pensieve, a Mac app that indexes your files, repos and past agent sessions locally and hands the agent a few passages through MCP. One HN commenter says they have seen about 20 tools like it and a new one arrives every month. THE THING The site claims a benchmark with Claude Sonnet in Claude Code across four repos of 504 to 34,847 files: 2.0 tool calls per lookup instead of 3.4, $0.041 instead of $0.050, and the right file found in 131 of 144 runs against 123 of 144 with grep. It states that a single grep still wins on exact error messages. The scripts are said to be open source in bench/ so you can rerun them. Author numbers, not run by us. Apple Silicon, macOS 14 or later. ✂️ STEAL IT Steal the benchmark shape: same tasks, with and without the tool, graded on whether the right file was found, and check the cost per lookup, not just the speed. Run it on your own repo. Try today VERDICT Only if you live in a big repo on a Mac and already feel the grep loop. Skip it if your repos are small or your questions are exact strings, where the author’s own data says grep keeps up. Pensieve site · Repo · HN discussion Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes. More like thisLess like thisDig deeperAlready knew it |
04 131 stars · MIT · repo created Oct 10 · HN: 47 points · image: project banner 📞 Your agents, on a conference callAgent Standup reads the session files you already have. SCENE A developer publishing as Atoll92 got tired of not knowing which of several Claude Code sessions was stuck. The README says Agent Standup turns the session files Claude Code and Codex already write into a live team call, with sub-agents as tiles at desks. THE THING It reads local session files and, per the README, never calls Claude unless you turn on translation or jokes. Tiles show a token counter per agent, an amber hourglass when a command has run over a minute, a sleep icon after three quiet minutes, and a failed-call counter. It runs as a VS Code extension, a terminal command or a desktop app. The README says the try-it route is npx @double_geste/agent-standup --open with the built-in demo, no data of yours. It is marked unofficial and unaffiliated with Anthropic or OpenAI. Not run by us. ✂️ STEAL IT Take the signals, not the cartoon: per-agent token count, a "running over a minute" flag and a "quiet for three minutes" flag are what you want when five agents run at once. Watch VERDICT Playful and cheap to try on the demo. The stuck-agent flags are the useful part; whether they are accurate across real sessions is untested by us. Repo + README · Project site · HN discussion Tap one, then press send in the email that opens. Researched replies usually land within about 10 minutes. More like thisLess like thisDig deeperAlready knew it |
📊 THE ATTENTION GAP HN points and GitHub stars at 06:25 IDT. Attention, not quality. | Sarnacki rewrite post | | 332 HN | | Chrome Relay | | 253 stars | | Agent Standup | | 131 stars | | Pensieve | | 37 HN |
|
🗂️ THE BORING STUFF Oct 9 · agent-browser 0.39.0. Adds an experimental Obscura engine behind --engine obscura, which the release notes say still has accessibility and rendering gaps, and private CA trust for CLI HTTPS requests with --ca-cert and --use-system-ca for networks that inspect TLS. Release notes → Oct 11 · Ollama 0.40.3. ollama launch claude and ollama launch codex no longer show the connectors and service-tier warnings, and a background model upgrade for embedding and other request performance was paused. Release notes → Official release pages checked this morning. Nothing installed or run. |
🧪 ONE THING TO TRY TONIGHT List the coin flips before the prompt. Before you ask an agent for a rewrite, write five lines: what must stay compatible, what can be dropped, memory limit, what happens to data on a crash, latency or throughput. Paste them in with the prompt. |
What this issue could not see. HN, GitHub, official release pages and first-person posts were read; projects were not run. Deduped against every published issue. Reddit returned rate-limit errors earlier in the week; X, Product Hunt and Bluesky returned nothing usable and LinkedIn was a login wall, so none of those are covered today. Video discovery found nothing current that passed review, so no video is recommended. |
|
| Images: credited in each report; source screenshots and announcement artwork, not our tests. |
|