How it works
Roundtable has two modes. Both fan your question out to the configured heads in parallel; they differ in whether the heads get to see and respond to each other.
Advisory (default)
One round. Each head answers your question blind — it never sees the others' responses. You get every independent perspective side by side, plus a synthesized recommendation at the end.
Use it when you want breadth: diverse, uncorrelated takes on a decision, fast.
roundtable ask "What's the biggest risk in this rollout plan?"
┌──────────────────────────────┐
│ your question │
└──────────────────────────────┘
│ │ │ │ │
▼ ▼ ▼ ▼ ▼ (parallel, blind)
🤖 🧠 🟣 🟠 🔵 ✨
│ │ │ │ │
└─────┴──┬──┴─────┴─────┘
▼
synthesized answer
Deliberation
Multiple rounds. After round one, each head sees the whole table — every other head's answer — and revises its own. A Claude chair reads the table at the end of each round and prints a verdict:
VERDICT: CONSENSUS— the heads have converged; the council stops early.VERDICT: CONTINUE— meaningful disagreement remains; run another round.
Use it when the decision is high-stakes and you want the models to actually reconcile their disagreements rather than just vote.
roundtable ask "Postgres or SQLite for this workload?" --rounds 3
round 1 each head answers blind
│
▼
chair reads the table ─► CONTINUE
│
▼
round 2 each head sees ALL answers, revises
│
▼
chair reads the table ─► CONSENSUS ✓ stop early
│
▼
synthesized verdict
The chair
The chair is a dedicated Claude seat that doesn't compete — it adjudicates. Each round
it summarizes where the table agrees, where it splits, and decides whether another
round is worth running. --rounds N is an upper bound; consensus ends it sooner.
Research mode
Pass --research to let heads that support tool use pull in fresh context before they
answer. Off by default — keeps runs fast and deterministic.
See the CLI reference for every flag, and The heads for who sits at the table.