Agent Committee — Development skill for Claude Code

Committee – a decision committee of local agent CLIs

Runs Claude Code, Codex, Grok and Antigravity (agy) as one committee on one problem. Each agent gets a written role, answers blind, then cross-examines the others' anonymized answers, then a chair writes the decision memo. One Python file, standard library only, no API keys – each CLI uses its own login.

problem.md ─┬─> claude (Lead analyst)    ─┐
            ├─> codex  (Operator)        ─┤  round 1: blind, parallel
            ├─> grok   (Devil's advocate)─┤
            └─> agy    (Risk officer)    ─┘
                         │ answers shuffled into Member A–D, vendor names hidden
                         v
            each member attacks / revises / holds, votes again   round 2..N
                         v
            chair (claude by default, falls back to next member) writes decision.md

Setup (10 minutes)

1. Python 3.9+ (python3 --version). Nothing to pip install.
2. Log in to each CLI once, interactively, so headless runs can reuse the login: claude · codex login · grok login · agy (sign in on first launch).
3. Put this folder anywhere, e.g. ~/committee, and run:

cd ~/committee
python3 committee.py --check

Nothing is installed outside this folder, and there's no Grok config file to set up – see "Grok sandbox".

`--check` pings each CLI headless and prints OK/FAIL with the reason. Fix any FAIL before a real run (see troubleshooting). Two working members is the minimum.

Run it

# quick question, free-form votes
python3 committee.py "What should our Q1 partner recruitment motion be?"

# a decision with fixed options – votes are tallied against them
python3 committee.py -f problem.md --options "Build,Buy,Wait"

# give it material: short files are pasted into the brief, a folder is readable by the agents
python3 committee.py -f problem.md --context pricing.md --context notes.md --cwd ~/code/repo

# subset, more rounds, different chair, longer answers
python3 commi