用 Jev 给 Claude Code 多智能体调试团队做分支决策

delost @thedelost原推发布于 2026.10.07

在 X 上打开原推

I don't understand why everyone isn't doing this yet. Anthropic's own Claude Code docs show how to debug with a team of Claudes that try to prove each other wrong, while Opus 5.5 only reads the verdict the whole idea: competing hypotheses one agent finds one plausible cause and stops looking. five teammates each get their own theory and try to kill the others' - the one that survives is the real cause the lead: Opus 5.5 on high, writes the hypotheses, never digs itself the investigators: Sonnet 5.5, five of them, one theory each, read-only until the plan is in the skeptic: Fable 5.1, owns no theory, only shows up at three points: → when a theory claims proof: which log line shows it? → when two theories agree: is that evidence or an echo? → before the fix ships: does it explain every symptom? Sonnet 5.5 digs. Fable 5.1 doubts. Opus 5.5 calls it Jev engineering is the same move one layer down: the forks inside each investigation (which log, which file, dig or drop) go to Jev in under half a second, and the debate only burns tokens on the ones that split turn it on, then start the debate in plain English: "Spawn 5 teammates to investigate different hypotheses. Have them try to disprove each other's theories, like a scientific debate" - the full team > Opus 5.5 on high leads from plan mode, so teammates investigate read-only first > 5 investigators on Sonnet 5.5, one hypothesis each > skeptic on Fable 5.1, owns nothing, attacks every proof > findings doc updated only with what survives > split panes with --teammate-mode auto, Ctrl+T for the task list paste the team and this prompt into Claude Code ↓ "Set up a debugging team for this repo: 1. In ~/.claude/settings.json add CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 under env, set effortLevel to high and teammateMode to auto 2. In ~/.claude/agents, draft two subagent definitions > investigator with model: sonnet and read-only tools, whose job is proving one hypothesis and disproving the others with logs, not opinions > skeptic with model: fable and read-only tools, whose only job is attacking proof claims, agreements and the final fix > Skip any that already exist and list them 3. Add a TeammateIdle hook that keeps a teammate working (exit 2) while its hypothesis has no evidence either way, and create docs/findings.md where only surviving theories get written 4. Find anything that would override this (CLAUDE_CODE_SUBAGENT_MODEL, CLAUDE_CODE_SUBAGENT_MODEL_FORCE, CLAUDE_CODE_EFFORT_LEVEL). Report it, change nothing Show me every change as a diff first. No edits until I say go." ↳ https://code.claude.com/docs/en/agent-teams

这条是引用着另一条写的：x.com/hanakoxbt/status/2101709924828934222