Claude Code 组队：Jev 半秒内选文件和工具

rvaniaaa @rvaniaaaa原推发布于 2026.10.06

在 X 上打开原推

I still don't understand why people hand one agent every task when you can build a team with no weak spots the team: → Critic, Fable 5.1: checks in at three points: where modules meet, after a test fails twice, and before the PR goes out → Architect, Opus 5.5 on high effort: owns the plan and the final merge → Developers, Sonnet 5.5 on medium effort: UI/UX and backend, each owning its own files so they don't overwrite each other → Jev: picks which file to open and which tool to run in under half a second, and a plain script does the rest teammates share one task list and claim work with file locking, built in one prompt sets up the config and shows every diff before it touches anything: "Configure my Claude Code workspace for agent teams: In ~/.claude/settings.json, add "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" under env. Set high effort for the main session inside modelSettings for claude-opus-5-5, since a top-level effortLevel does not apply to Opus 5.5. In ~/.claude/agents, create any missing subagent definitions: a UI/UX developer and a backend developer on Sonnet at medium effort, and an adversary on Fable. In CLAUDE.md, add a rule: the adversary reviews at contract boundaries, when the same test fails twice, and before any PR is opened. Show every config diff before applying anything." then tell Claude: "Spawn three teammates: UI/UX and backend developers, plus the adversary." http://code.claude.com/docs/en/agent-teams

这条是引用着另一条写的：x.com/rvaniaaaa/status/2101709732314558639