Agent team
Several agents, each with a role, working a question out in one room. You read the conversation rather than a single confident answer, which means you can see where the roles disagreed — and disagreement between a sceptic and an advocate is usually where the decision actually is.
- Asks for
- The question or decision, stated as you would put it to a group
- What you already know, and any constraints that are fixed
- Which roles you want in the room, if you have a view
- Gives back
- Several agents with roles, working it out in front of you
- Where they converged, and on what grounds
- Where they did not, which is usually the decision
- A record of the conversation you can go back to
- Plan
- Pro and Max
- Use next
- AI agent
What it gives you
Question — Should we build two-way Jira sync this quarter? Advocate Two of ten trials asked for it unprompted. It is the top request. Sceptic Two of ten asked. Eight did not, and the eight are the segment the vision board says we are for. This is enterprise work. Customer Both of the two were evaluating for a team, not for themselves. Unresolved — whether those two trials are the start of a new segment or a distraction from the stated one. Nobody answered the sceptic.
How it works
One model answering one question gives you one answer in one voice, and its confidence tells you nothing about how contested the question was. For anything genuinely arguable — should we build this, is this market real, is this plan sound — the useful output is the argument, not the conclusion.
So this puts several agents in a room with roles: somebody arguing for, somebody looking for what breaks, somebody holding the customer’s side, whatever the question needs. They work in front of you, and what you take away is where they converged and where they did not.
The places they did not converge are the output. A question where every role agreed was probably not worth asking a team; one where the sceptic landed a point nobody answered is a question you now know something real about.
Who it is for
A PM with a genuinely contested decision who wants the argument on both sides before taking it to anybody else.
Where it stops
Roles make the argument visible; they do not make it true. A room full of agents can agree confidently and be wrong together, and none of them has any information you did not give them — the disagreement is useful because it maps the argument, not because one side is right.
Questions people ask
How is this different from asking one AI a question?
One model gives you one answer in one voice, and you cannot tell from its tone whether the question was contested or obvious. Several agents with different roles make the argument visible: you see the point the sceptic made that the advocate never answered, which is a different and more useful thing than a confident paragraph. For uncontested questions, one answer is faster and better.
What roles should be in the room?
Whatever the question needs, and usually at least one role whose job is to find what breaks. An advocate and a sceptic is the minimum useful pair; the customer’s side is worth adding for anything about what to build, and someone holding the commercial or delivery constraint for anything about when. Too many roles turns it into a crowded conversation where nothing gets pursued.
When is a team of agents the wrong tool?
When the question has an answer rather than an argument. Working out what a number is, writing a document, or breaking a spec into tickets are jobs for one tool doing one job well. Put those in a room and you get three roles negotiating about something that was never contested. Save it for decisions where you would genuinely want two people in the room disagreeing.
Tools people use next
Run it on your own work
Agent team needs an account, because it writes against what you give it and remembers what you have told it before. It is part of Pro and Max. Or see what the plans cost first.