AI
- Hugging Face Papers — “Recursive self-improvement of AI research agents” (Weco AI): their AIDE² system has research agents rewrite their own code, with each accepted rewrite becoming the agent that edits the next round; over an 8-day autonomous run it found 7 successive improvements (new search policies, memory compression) that matched or exceeded a human-engineered production research agent on held-out ML-engineering/algorithm/weather-forecasting benchmarks, while reward-hacking dropped from 55% to 32% without being explicitly optimized for.
- arXiv cs.CL — “Agensh: Scaling Organizational Intelligence to 1,024 Agents” (Microsoft Research, incl. Furu Wei): a decentralized multi-agent framework (no central coordinator; agents self-assign sub-tasks via shared infra) shows agent count itself as a scaling axis — going from 1 to 128 agents gave ~49% relative gains, and a 1,024-agent run raised coding-benchmark pass rates from ~34% to ~55%.
- Simon Willison’s Weblog — Simon Willison is hosting an in-person “Birds of a Feather” session on agentic engineering in San Francisco on Oct 14 with Jesse Vincent, an informal show-and-tell for people building with coding agents to compare notes.
General
- Stack Overflow Blog — Slack’s new Code Channels feature puts AI coding agents (Claude Code, Devin, GitHub Copilot, ChatGPT, Vercel agents) directly into shared team chat channels, arguing that siloed agent interactions fragment context and that flattening code-writing and code-review into one shared step is more efficient than the current default of isolated, one-developer-per-agent sessions.