In an Oct 21 engineering post, GitHub says it's changing its write quorum from the whole repository to individual references, so agents and developers pushing to their own branches no longer queue behind each other, which removes the 25ms round-trip penalty that kept its busiest repos out of Azure. GitHub says pushes grew 4.9x year over year to 3.35 billion a month (earlier post).
Since Oct 20, the first interactive V3 session in a folder shows a "Trust this workspace?" screen before Kiro CLI loads that folder's hooks, agents, MCP servers, steering, workflows, or language server config, so a cloned repo can't change agent behavior until you say so. Answers are saved per workspace root under /.kiro/workspace roots/ , though your home directory is never remembered and asks every launch…
Announced Oct 21 in public beta for Claude Enterprise, it runs a deep baseline scan across projects of up to eight GitHub or GitLab repos, then rescans changes as commits land, marking findings fixed when the fix ships and reopening them if they come back. Findings are also readable through a new API, an MCP connector for Claude and Claude Code, and webhooks, with $25K/month in free credits for two months through Dec 31 (Help Center)…
Announced Oct 21 as a research preview for US Pro and Max subscribers, Duet runs on Claude Opus 5.5 in its own VM with a browser, checks your connected apps every few hours, sends a morning brief, and asks for approval before it sends, pays, signs, or deletes anything. It's a separate app on iOS, Android, and claude.ai, and its background checks count against your plan's usage limits…
Released Oct 21 on Hugging Face, Flint is a 40B mixture of experts model with 2.5B active parameters that Recursive says hit about 131 tokens/sec on an M4 Pro MacBook Pro and beat other sub 50B models on SWE Bench Verified (its own numbers). The company says its Eureka Machine research agent helped with nearly every step, and warns Flint needs big tasks broken into smaller steps…
Released Oct 21, new agent sessions now run on the Copilot SDK based harness, and a preview Compare Models mode in the Agents window runs one prompt across Copilot, Claude, and Codex harnesses with an optional judge model to score and synthesize results. Experimental on device MAI Code 1.1 Flash skips AI credits but needs an NVIDIA GPU with at least 96 GB of memory (Visual Studio Magazine)…
Spotted something we missed? Start a thread.
