Council answers
A council runs several models, or several agents, on the same question and merges their answers into one. Use it for decisions that matter, where a second and third opinion are worth the extra time and credits.
How a council works
Every seat answers the question. In app chat, the API, MCP and Slack, each seat retrieves within its own agent's knowledge scope. On the website widget, every seat works from the widget's single search, so no seat can widen what a visitor sees.
If the seats disagree, one debate round runs so they can reconsider together.
A lead seat reconciles the views and writes the final answer. In chat, the lead is the chat's current model using the current agent.
While it runs you'll see progress such as "Running Council across N seats…", "The members took different views — reconsidering together…" and "Synthesizing the council's answer…". A seat that errors is retried once; a seat that fails or times out is left out and the council answers from the seats that finished. If the final merge cannot complete in time, you get the strongest single response, marked as such.
After the answer, a card titled How the council answered shows each member, its stance (agrees, partly agrees, dissents), a lead tag, and a debated chip when a debate round ran. Click a member to see its initial and final responses.
Seats and limits
A council needs at least two members and can hold up to nine. In chat, adding a tenth removes the oldest member.
In chat, members can be models or agents. An agent seat answers with that agent's persona and, outside the website widget, its knowledge scope. A council you create in the agent builder is made of models.
If a member's model or agent is later removed, retired, turned off, or no longer accessible to the person who set up the council, that seat is set aside and that person gets a notice such as 'Your council "X" no longer includes Y'. A council left with fewer than two seats drops out of the picker until access is restored.
Where to turn it on
For one conversation in chat
In the composer, open the model button next to Send (titled "Model, thinking effort & answer mode").
In the Council row, turn Council on, then choose members under Add members from the Agents and Models lists, or pick a saved preset under Presets.
Send your question. If fewer than two members are picked you'll see "Select at least two models or agents for Council, or turn Council off."
You can Save the selection as a preset, and later Rename, Update to current or Delete it. Presets belong to the whole organization, and any member can create and manage them.
As an agent's default
In the agent builder's Behavior step, under Model, choose Council instead of Single model and pick a saved council. Click New council to create one in place: give it a Council name, pick two or more models, and click Create council. Save the agent first; a new agent needs its permanent handle before a council can anchor to it.
Once set, every answer that agent gives in app chat, the API, MCP, Slack and email runs the council, unless the person asking picks a different model or council for that turn. On Slack, --council=<name> picks a council per question. The website widget runs the agent's council too.
Time and cost
A council is slower than a single model. Every seat has to answer before the merge, and a debate round adds another pass. Expect several minutes on a thorough question; the seat phase alone is allowed up to about nine minutes before stragglers are dropped.
There is no fixed multiplier. Every seat's usage plus the final merge is metered and billed at the normal rate, so a council of three costs roughly three times a single answer, and more if a debate round runs. The builder puts it plainly: "about N× the cost and latency."
Before the answer, a temporary credit hold is placed for the chosen depth multiplied by the number of seats. If a council fails after work has started, only the usage that actually happened is charged.
See Credits and the rate card for how metered usage is priced.
Plan availability
Council is included on every current plan, and its usage comes out of the same credit balance as any other answer. If your organization has a custom plan without it, a council answer on a signed-in surface returns "Council answering is not included in your organization's plan. Ask your administrator about your plan." On the website widget the agent quietly answers with a single model instead.
When to use it
Good fits: policy questions, pricing and contract interpretation, anything a person will act on without double-checking, and questions where your sources might disagree. Poor fits: quick lookups, conversational turns, and customer channels where a fast reply matters more than a second opinion.