← ALL POSTS
Pattern6 MIN READ

The AI Roundtable: A Pattern for High-Stakes Decisions

An AI roundtable stages multiple frontier models around one hard question. Here's the pattern, when to run one, and how to read the output.

Call it a roundtable when the question is too important for one model to answer alone. Six frontier LLMs sit around the same prompt. Each opens with a position. Each attacks the weakest opening. After a few rounds, a synthesis pass writes the verdict.

Why it's shaped like a meeting

Because the failure mode is the same. In a meeting with one person, you get one opinion. In a meeting with six people where no one's allowed to disagree, you get one opinion. The value of a roundtable is the disagreement, and the discipline is to keep it visible in the output instead of averaging it away.

What to bring to an AI roundtable

  • A real question — not a lookup.
  • The constraints. Budget, deadline, non-negotiables.
  • Any evidence you already have. Documents, prior analysis, links.
  • The decision you'll make with the output. "So what?" should be answerable.

How to read the output

  • Read the verdict first. Then the dissent.
  • If the dissent is stronger than the verdict, trust the dissent.
  • If every model agreed, be suspicious — you may have asked a question that only had one answer.
  • If the models cited sources, click through. A verdict without sources is a vibe.
// TRY IT WITH YOUR QUESTION

Example prompt to paste in the composer:

Convene an AI roundtable on [my decision]. Give me the verdict, the dissent, and the sources.
Stage a roundtable