When the agent edits or creates a file in a subdirectory, it now picks up that directory's AGENTS.md and nested always-apply rules even if it never read a file there. Custom subagents pinned to a blocked or unavailable model now fall back to the parent agent's model instead of erroring, and agent env deploy --env can redeploy onto an existing Cloud Agent environment.
Agents open, write and land tasks through MCP tools, and each landing is tested against the exact current head of main and replayed if main moved, while still writing ordinary Git commits so GitHub and deploys keep working. In Murmell's own benchmark with 8 Claude Code agents, main went red 0 of 102 times versus 14 of 104 under Git; try it with avp up in a checkout (code)…
Shown on HN Oct 25, the Apache 2.0 Rust tool has AI workers read a site's docs, frontend code and live network traffic to generate read and write actions, signing requests with credentials kept in your OS keyring so the model never sees them. Writes stay off per site until you run hycli writes SITE allow , and you hook it into Claude Code, Cursor or Codex with "command": "hycli", "args": ["mcp"]…
Shown on HN Oct 25, the MIT licensed app (formerly Triage) ranks what needs you across GitHub, Slack and MCP connected tools, then fixes CI, addresses review nits and drafts replies using your own Claude Code or Codex login, with every push, PR, comment and message held for your approval. It's macOS only, installs via a curl script because the app isn't notarized yet, and stops when your Mac sleeps…
Accounts are moving one at a time: each playbook becomes a skill with the same instructions and scope on the Customize page, Settings Playbooks goes read only, and Devin's System playbooks arrive as a built in System Skills plugin you can't uninstall. Existing !macro shortcuts, schedules, automations, Jira and Linear mappings, and the Playbooks API keep working by running or updating the converted skill, while new prompts can call one with @skills:name…
A meta agent reads a task agent's trajectories and writes reusable interfaces (bulk spreadsheet edits, in tab browser requests, app APIs), so Qwen3.8 27B hits 21.3% on OSWorld V2 versus 5.6% with fixed GUI tools and 20.6% for Claude Opus 4.8, at about $1.10 a task versus $72 (self reported). Code is at simular ai/Agent S…
Spotted something we missed? Start a thread.
