Round Tables · structured disagreement
Three specialists weigh in. You arbitrate.
Asking an AI for three options is the most common move in the world, and it almost never works. You get three variations written by the same mind, ranked by the same priorities, phrased to be equally appealing. That is not a choice. That is a menu designed by someone who does not care which one you pick.
A Round Table produces a real disagreement because the positions are assigned before anyone speaks. Someone argues for growth. Someone argues for margin. Someone argues that the question is wrong. Then one of them decides, in writing, and says what it rejected and why.
What an AI agent orchestration platform actually has to do
Most platforms in this category orchestrate by sequence. Agent A runs, agent B runs on A's output, agent C formats the result. It is a pipeline, and it has a structural weakness: the second agent cannot tell the first one that the premise was wrong. It can only build on it.
Real orchestration requires three things a pipeline does not have. Positions that are genuinely opposed rather than sequential. A mechanism for one agent to reject another's contribution outright. And an arbitration step that produces a decision rather than an average.
Round Tables are Crew Studio's answer to the third one specifically. Crew Chats handle the open debate, which is right for production work. A Round Table handles a question that has a real trade off inside it, where the value is not in producing more but in deciding better.
The test for an AI agent orchestration platform is therefore not how many specialists it can run at once. It is whether two of them are allowed to reach opposite conclusions, and whether anything in the system is built to settle that rather than average it.
How a Round Table runs
- You state the question, not the task
- The input is a question with a genuine tension in it. Should we raise the price or raise the scope. Do we launch in September or wait for Q1. Is this a positioning problem or a channel problem. A Round Table convened on a task will resolve immediately and tell you nothing.
- Positions are assigned before anyone writes
- Crew Studio picks three to five specialists whose roles pull in different directions, and gives each one a position to defend. This assignment happens first, on purpose. A specialist that discovers its position while writing tends to end up wherever the argument was easiest.
- One specialist arbitrates, and it is not you
- The arbitration is done inside the Round Table, by a specialist whose role is to decide. It states what it chose, what it rejected, and the reason for each rejection. You get a decision with a rationale rather than three options and a shrug.
- You keep the final call
- The arbitration is a recommendation, not an execution. Nothing ships from a Round Table. You read the synthesis, you agree or you do not, and the rejected options remain available if you want to reopen one.
WATCH IT RUN
Watch a Round Table resolve
A real Round Table from a Studio running a paid newsletter. The question had a genuine trade off, and the arbitration went against the position most people would have expected.
Notice what the arbitration does that a list of options does not. It rejects Position 2 by reading the actual cancellation quotes rather than by preferring a different theory. And it sets a review date with a specific metric, which means the decision can be proven wrong later.
The question
Churn is at 6 percent monthly. Do we add features, drop the price, or narrow the audience?
Position 1 · Growth
Narrow the audience. Six percent on a paid newsletter is an acquisition problem wearing a retention costume. We are selling to people who were never going to stay.
Position 2 · Product
Add depth, not features. The cancellation reasons cluster on "not enough for the price", which is a perceived value gap, not a feature gap.
Position 3 · Pricing
Do not drop the price. A lower price on the same product accelerates the same churn and removes the room to fix it.
Sofia · Arbitration
Narrow the audience. Rejecting Position 2: the cancellation quotes say "not for me" more often than "not enough", which is targeting, not depth. Rejecting the price drop for the reason Pricing gave. Revisit in 60 days on cohort retention, not on total churn.
4 positions · 1 arbitration · 2 rejections, each with a stated reason
How the positions get assigned
The Crew is read before the question is.
Crew Studio looks at which specialists exist in the Studio and what their roles actually optimise for. A Crew where every specialist optimises for the same thing cannot produce a Round Table worth reading, and the platform will say so rather than staging one.
Positions come from role conflict, not from randomness.
A pricing specialist and a growth specialist disagree structurally, because one protects margin and the other protects volume. That conflict is real and it is where the assignment comes from. Handing a specialist a position it has no reason to hold produces an argument nobody believes, including the specialist making it.
One position is always allowed to reject the question.
A meaningful share of decisions that feel stuck are stuck because the question is wrong. One seat at every Round Table can say so, and it is used more often than most operators expect. In the example above, Growth effectively did this by reframing a retention question as an acquisition one.
The arbitrator is chosen for distance, not seniority.
The specialist that decides is the one with the least stake in the outcome. A pricing question is not arbitrated by the pricing specialist, for the same reason you do not let one side referee.
Assignment is the step that gets skipped, because it looks like setup rather than substance. It is also the step that decides whether an AI agent orchestration platform produces a debate or a chorus, and nothing downstream repairs a table where everyone was pointed the same way.
What a Round Table leaves behind
The decision, with its reasoning.
Not a conclusion but the path to it, including which evidence carried weight. Three months later this is the difference between reopening a decision cheaply and relitigating it from scratch.
The rejections, each with a stated reason.
This is the part operators underestimate. A rejected option with a reason attached is a decision you can reverse intelligently when the reason stops holding. A rejected option with no record is just a thing you vaguely remember not doing.
A review trigger.
Every arbitration sets a condition under which the decision should be looked at again, tied to a named metric rather than a date. A decision that cannot be proven wrong is not a decision, it is a preference.
Context for the next Round Table.
The record sits in the Studio, so the next debate on an adjacent question starts from what was already settled. Crews stop relitigating positioning every time a pricing question comes up.
Why arbitration beats consensus
Ask a group of AI agents to agree and they will, quickly, because agreement is the path of least resistance in a system trained on agreeable text. The result is a position nobody would have argued for on its own. It reads reasonable and commits to nothing.
Arbitration produces the opposite. Someone has to lose, in writing, with a reason. That single constraint changes the quality of the debate upstream: a specialist that knows its position can be rejected argues it properly rather than hedging toward the middle.
It also leaves a record worth having three months later. When the decision turns out to be wrong, you can read what was rejected and why, rather than reconstructing your own reasoning from memory. That record lives in the Studio and outlasts the conversation.
This is where an AI agent orchestration platform is usually thinnest. Running several specialists at once is an engineering problem, and it is largely solved. Deciding between them is an editorial one, and it has to be designed rather than scaled.
Who Round Tables are for
Operators who decide alone.
There is nobody to disagree with you, so you have been arguing both sides in your own head and calling it a decision. A Round Table externalises the argument so you can read it instead of holding it.
Consultants who need to show the reasoning.
A client rarely wants the answer without the path to it. A Round Table produces both, and the rejected options are usually more persuasive than the chosen one.
Anyone stuck between two options for more than a week.
If a decision has been open that long, the missing thing is not information. It is someone willing to reject one side out loud.
Operators who have to explain a reversal.
Changing direction is easy. Explaining to a client, a partner or your future self why the previous direction made sense at the time is harder, and it is what the record of rejections gives you.
WHERE IT DOES NOT HELP
When you do not need a Round Table
If the question has one right answer, skip it.
A Round Table is built for trade offs, not for facts. Asking three specialists to agree on a date or a figure wastes four minutes to confirm something one of them already knew.
If you have already decided, skip it.
A Round Table convened to validate a decision you have made will find arguments for it, and you will have paid four minutes for reassurance rather than a second opinion.
If the brief is a production task, use a Crew Chat instead.
Round Tables resolve a question. Crew Chats produce work. Convening a debate to write a newsletter is the wrong tool applied carefully.
If nobody in the Crew has a stake, the debate is theatre.
Three specialists with the same brief and no conflicting priorities will converge immediately. Round Tables need roles that genuinely pull in different directions.
SIDE BY SIDE
Round Tables compared to asking one model twice
Three rows go the other way, and on a factual question the gap is not marginal. Asking one model twice is faster, cheaper and entirely sufficient when there is nothing to trade off.
The honest framing is that these two tools answer different questions. Asking a model twice answers «what could this look like». A Round Table answers «which of these should we do, and what are we giving up». Most operators need the first one far more often than the second, and reaching for a Round Table on every question is how the feature stops being useful.
| Capability | Round Table | Asking one model twice |
|---|---|---|
| Positions argued | 3 to 5, assigned in advance | 2 answers from the same model |
| Independence between positions | Roles with different priorities | None. Same weights, same bias. |
| Arbitration | A specialist decides and states why | You decide, without a rationale |
| Record of what was rejected | Kept in the Studio | Lost when the tab closes |
| Review trigger set | Yes, with a named metric | No |
| Time to result | Minutes | Seconds |
| Cost | Higher | Lower |
| Usefulness on a factual question | Low. Overkill. | High. Direct. |
| Usefulness on a trade off | High | Low |
KEEP READING
Related
QUESTIONS
Questions about Round Tables
What is a Round Table?
A Round Table convenes three to five specialists on one question with structured roles. Each argues a position, one arbitrates, and Crew Studio returns a single synthesis with the reasoning behind it. The operator reviews one decision instead of forty messages.
Do I have to review everything they produce?
No. Crew Studio returns a synthesis: three ranked options, the reasoning for each, and a recommended path. The full thread stays available but is not the deliverable. Reviewing a synthesis takes about twenty seconds instead of twenty minutes.
Can a Round Table end without a decision?
Yes, and that is a legitimate outcome rather than a failure. When three specialists hold their positions and none concedes, the disagreement is the finding: it usually means the question was underspecified or the trade off is real and yours to make. What you get back is the map of the disagreement, not a manufactured consensus.