A model that answers in numbers, OpenAI's one-hour counterpunch, and instruction files that quietly rot

Your agent's instruction files are quietly rotting. Addy Osmani's new post argues that the config files coding agents read — CLAUDE.md, AGENTS.md, skill definitions — need regular audits, because rules written for last year's models turn into stale junk that can actively hurt this year's ones. Claude Code's /doctor command, run inside a session, is the built-in way to check. The instinct on most teams is to accumulate agent instructions; the evidence points the other way, and if you advise teams on agent delivery this is the argument to bring: prune on a schedule, and test whether each installed skill still earns its keep. An audit of 100 well-known open-source repos in June turned up lint-type config junk in 62% of them, needless context in 42%, and skill definitions leaking where they shouldn't in 35% — while Anthropic stripped over 80% of Claude Code's own system prompt for its current model generation and saw nothing get worse on its internal coding tests. addyosmani.com

“Your coding agent's configuration has a half-life.”

ChatGPT's mobile app now takes voice orders for agent work. OpenAI added voice-triggered workflows to the ChatGPT mobile app: you speak, and it drafts documents, writes emails or summarizes Slack threads from the Work tab, with the conversation pickable later on desktop. Voice plus a workspace tab is OpenAI's push to be the place where agentic work starts — a deliberate contrast with Anthropic, which merged its chat and workspace products together rather than keeping the two separate. The split runs by tier: paying Plus and Pro users get the document and Slack workflows, while free and Go users get plugins and connected apps. TechCrunch

A playground for Google's new talking models. Google shipped Gemini 3.8 Flash TTS and Flash-Lite TTS, its most expressive text-to-speech models yet, and Simon Willison immediately built a free, bring-your-own-API-key playground for trying them — including multi-speaker audio you can save to a bookmarkable URL. Voice is the next interface clients will ask about: these models design voices from plain-text descriptions, clone a voice from a 30-second sample, and ship with a library of more than 2,000 production-ready voices. Generating a minute and eighteen seconds of audio took about 20 seconds and cost 2.74 cents. simonwillison.net

OpenAI counter-launched within the hour, and a price war is on. Anthropic and OpenAI released new models within about an hour of each other this week, and the real story isn't capability — it's the price cuts that came with them. Simon Willison's comparison piece is the useful one-stop read. Cheap inference keeps changing what's economic to run: Opus 5.5's cache-read price — the cost of re-feeding an agent its own conversation — dropped 60%, and agent workflows live or die on cached context. GPT-6 Luna lands at $0.10 per million input tokens and $0.50 per million output, half what its predecessor tier cost, and Willison has switched his own defaults in Codex and Claude Code to GPT-6 Sol and Opus 5.5. simonwillison.net

Anthropic's biology lab says Claude found a CRISPR-like system. Anthropic confirmed it runs a real wet lab in the Bay Area, and announced that Claude found a previously unknown enzyme system hidden in bacteriophage DNA — one whose repeating structure echoes CRISPR. The company has named it array-associated reverse transcriptases. It's the clearest public demo yet of the agentic playbook applied to science: the model did the searching, human scientists did the pipetting, and the run is agent orchestration at a scale worth studying whatever you think of the biology. The search took about 21 hours of roughly 950 agents burning through 210 million tokens; the lab works only at the two lowest biosafety levels, and Dario Amodei acknowledged a Stanford team had previously found something similar. TechCrunch

The EU AI Act has rules specifically for agents — here's the short version. Angie Jones has written a practical walkthrough of how the EU AI Act treats AI agents: legal duties attach to what the agent system is used for, separately from the model inside it, and interactive systems have had to disclose they are AI since August 2, 2026. This is the compliance question every European client will ask, and the answer is not 'the model provider handles it' — deployers carry logging and transparency duties, and giving the agent more autonomy doesn't hand those duties to the agent. High-risk systems need automatic logging that deployers generally keep for at least six months, and the Digital Omnibus pushed the high-risk deadlines out to December 2027 and August 2028. aaif.io

“The amount of autonomy can change which controls are reasonable, but autonomy does not move responsibility from the company to the agent.”

Bessemer raises $5.75B, all of it aimed at AI. Bessemer Venture Partners closed two new funds totalling $5.75 billion, dedicated across the AI stack — $1.75 billion for seed and early stage, $4 billion for growth companies. The money keeps signaling that AI-native companies are treated as a distinct asset class rather than a feature of software; Bessemer already counts more than 260 of them since 2022, including Anthropic, Cognition and Perplexity. The firm says it has put $3 billion into AI startups so far, and partner Byron Deeter calls the pace of scaling unlike anything Bessemer has previously backed. TechCrunch

“AI-native companies are scaling faster than any category of technology we've backed before”

Agentic engineering show-and-tell in SF on October 14. Simon Willison and Jesse Vincent are hosting an evening Birds of a Feather session in San Francisco for people building with coding agents to compare what they're trying and what they're learning. If you're in the Bay Area mid-October, this is exactly the room where the practitioner state of the art gets compared — explicitly framed as early explorations rather than product pitches, and nobody is required to present. It runs on Wednesday, October 14, and sharing what you're working on is encouraged but optional. simonwillison.net