The table and its seats

The table is where you assemble the panel that debates your topic. Each chair at the round table is a seat, and you fill seats with two kinds of AI panelists: the subagents you have saved in the AI Console, and one-time agents you create for this room only. This page explains the seating picker, how seat order sets who speaks first, the fields you set on a one-time agent, and how a seat appears once the council runs. The War Room is available on the Pro plan and above and is rolling out in waves, so an eligible account may not see it yet.

Before you begin

Use a Pro plan or above with the War Room switched on for your account. It is rolling out in waves, so an eligible account may not see it yet.
To seat a saved subagent, build one first in the AI Console. See Subagents and Create a subagent. You do not need a subagent library to start, since you can create one-time agents inside the room.

Quickstart

Open the seating dialog
In step 2, The Table, select Seat to open the dialog titled Seat the table.
Add panelists
Add saved subagents from the list, or select Add a one-time agent to create a panelist for this room only. Seat at least two and up to fifty.
Set the speaking order
The order you seat panelists is their speaking order in sequential rounds. Seat the voice you want to open the debate first.
Tune or remove a seat
Optional: override a saved seat's tools for this room, edit a one-time agent, or remove a seat you do not want.
Return to the rail
Select Done to close the dialog. Your lineup shows in The Table step, ready to start.

What it shows

The Table is the second step in the configuration rail, and the seating dialog behind it is where you build the council. The parts are described in the following sections.

The Table step

Step 2 is titled The Table with the prompt "Seat 2 to 50 panelists around the round table." Before you seat anyone it reads "No seats yet", inviting you to seat saved subagents or one-time agents. A Seat control opens the seating dialog. Once you have seated panelists the same control reads Edit, and the step lists your lineup in order.

The Seat the table picker

The seating dialog is titled Seat the table. A running count at the top shows how many seats are filled, for example "3 of 50 seated", with a status hint: Need at least 2, Up to 50 seats, or Room full. Below it sit your current lineup, an Add a one-time agent button, a search box, and a Done button. A tip in the dialog states that seat order is speaking order in sequential rounds. Each seat in the lineup shows its number and name, with controls to remove it, edit a one-time agent, or edit a saved seat's tools for this room.

Two kinds of seats

A seat can hold either of two kinds of panelist.

Saved subagent: one of the specialists you built in the AI Console. It keeps its own name, model, instructions, and tools, and you can override its tools for this room without changing the saved subagent.
One-time agent: a panelist you define here for this room only. It uses only the persona you type and the tools you allow, and it is never saved to your library. It disappears when you leave, unless you keep the lineup as a Saved lineup.

Inputs

The seating picker gives you these controls.

Add saved subagents: select a subagent from the list to seat it. The search box filters your subagents by name, description, or model, and each row shows the subagent's name and model.
Add a one-time agent: opens a form to create a panelist that lives only in this room (see the fields below).
Per-seat tools (saved seats): an optional override of which research tools that seat may use for this room only, without changing the saved subagent. The default is the subagent's own tool set.
Seat order: the order you seat panelists is their speaking order in sequential rounds. Seat your strongest voice first if you want it to open.
Remove: take a seat off the table.

The one-time agent form

Add a one-time agent opens a short form. Its fields, with their limits:

Name (up to 60 characters): the label shown on the seat, for example a role such as The Skeptic.
Provider: which family of models the model list offers. It is shown only when more than one family is available.
Model: an optional picker. Leave it unset and the room picks a cost-efficient default for you. Choose a model to fix the seat to it, and the platform's own GNG Analyst can also be chosen. Heavier models cost more credits per turn.
Reasoning: how hard the seat thinks per turn. A higher depth reasons more deeply and costs more.
Persona (up to 8,000 characters): free text describing how the panelist should think and argue. The form states this is the only instruction the agent sees, so be specific. A live character count is shown.
Tools: defaults to the full toolkit, so the seat can reach every research tool the room allows. Select Choose specific tools instead to narrow it to a grouped checklist, or Use all tools to return to the full set. A seat's tools always stay within your plan's tool access.

The save control reads Add to table when you create a seat and Save agent when you edit one.

Write a strong persona

For a one-time agent the persona is the whole brief: the panelist does not see your other chats, so it argues only from what you write here. A strong persona does four things.

Give the seat one clear point of view or role, distinct from the others at the table.
State its evidence standards: what it trusts, what it distrusts, and what would move it.
Say what it should challenge, so it presses on the weak points you care about.
Say how it should conclude, so its turn ends with a usable stance rather than a shrug.

Vague personas produce vague turns. The eight thousand character limit leaves room to define stance, priorities, and reasoning style.

Outputs and how to read them

Before you start, the lineup shows your seated panelists in speaking order, so you can confirm who opens and who follows. A reminder under the Start control tells you if you still need to seat between 2 and 50 panelists.

Once the council runs, each seat appears in the live thread as its own turn: a card headed by the seat number and the panelist's name, with a one-time agent marked as such and a subline showing the seat's model. While a seat speaks it carries a live status: Waiting, Thinking, Speaking, Done, or Failed. A short activity line names any research the seat reached for during its turn. For the full live view and the closing verdict, see The live room and the verdict.

How it is calculated

There is no scoring formula behind a seat. Each panelist is one voice in the debate, and the room meters each seat's turns in GNG credits as it speaks. What drives cost is the size and make-up of the table: more seats mean more voices per round, and a heavier per-seat model costs more per turn, which is why the picker warns you once you pass five seats. A one-time agent and a saved subagent cost the same for the same work; the only difference is whether the definition is saved. For how a room is metered and capped, see Spend cap and credits.

Data and timing

A panelist reads live market and fundamentals data through the tools you allow it, sourced from AlphaVantage (Nasdaq quotes delivered through AlphaVantage) and from EDGAR for public company filings, plus any of your own saved knowledge a seated subagent is bound to. The data is current to the platform's normal refresh schedule at the moment a seat looks it up.

Use cases

Seat three saved specialists you use often as a standing panel.
Add a one-time contrarian for a single debate without cluttering your library.
Restrict one seat's tools so it argues from fundamentals only.
Reorder seats so your strongest voice opens the sequential round.

Limitations and disclosures

One-time agents are not saved, so a table built from them is preserved only if you keep it as a Saved lineup or run it. The number of saved subagents you can keep is capped per tier, shown against your limit in the AI Console; one-time agents do not count against that cap. Every seat's tools stay within your plan's tool access. A council is a debate among AI panelists and a synthesized verdict; it is research support, not a recommendation, and it does not account for your personal circumstances.