Council / Multi-Agent Answers

What a Council is

A Council asks the same question to multiple AI models (participants) at once, then adjudicates their responses into a combined answer. Instead of trusting one model, you get several perspectives and a reconciled result — useful for high-stakes or ambiguous questions where one model alone might be confidently wrong.

How it runs

1. You pose a question to a council of chosen participants (models).
2. Each participant answers — this can happen over one or more rounds, where participants can react to the question (and, in multi-round setups, to the emerging discussion).
3. An adjudication step combines the participants' responses into a final answer.

The council, its participants, each round, and the adjudication are all recorded, so you can inspect what every model said and how the final answer was reached.

Billing — one record per participant

This is the key cost rule: every participant response is its own AI call, so it produces its own BillingRecord. A council with four participants over one round bills four calls; the adjudication step bills as its own call too. Cost scales with participants × rounds, and every call is itemized with a pricing snapshot.

When to use a Council

• Decisions where a wrong answer is expensive and you want cross-checking.
• Questions where models are known to disagree, and you want the disagreement surfaced.
• Evaluations where you want multiple independent takes before committing.

When NOT to use it

• Routine questions a single good model handles — a council multiplies cost for no benefit.
• High-volume automated calls where per-question cost matters; reserve councils for the questions that warrant the spend.

Tips

• Pick participants that are genuinely different (different providers/strengths) — three near-identical models add cost without adding diversity of opinion.
• Read the per-participant responses, not just the adjudicated answer — the disagreements are often the most useful signal.
• Keep an eye on participant count; cost is roughly linear in the number of participants and rounds.