🎓 delegation.school

Lessons / Power

The Decision Panel

For a genuinely two-sided decision — raise now or wait, build or buy, take the term sheet or counter, hire this exec or keep looking — don't ask your AI for one balanced answer. A single agent asked "should I do this?" hedges its way to the safe middle: "it depends, here are some pros and cons." That mush hides the very thing you need.

Instead, stand up a panel of advocates and make them argue. Spawn three independent agents over the same source material:

Each argues its side at full strength — no hedging. Then you (or a chair agent) weigh the arguments, name the single crux the decision turns on, and render a verdict with its reasoning. The disagreement between the bull and the bear is the signal a "balanced summary" would have averaged away.

This is the same multi-agent machinery as parallel subagents — with the opposite goal. There you wanted independent work and no conflict; here you want the conflict. It's how investment committees, war-gaming, and red-team/blue-team de-bias high-stakes calls. And it's distinct from a review pass: a review asks "is this artifact correct?" — the panel asks "what should we do?"

Try it now

Take a real decision you're actually sitting on and run it as a panel:

I have a real decision: [state it, and attach the context or docs]. Run it as a decision panel, not a balanced summary. Spawn three independent advocates: a BULL who builds the strongest case FOR, a BEAR who builds the strongest case AGAINST, and a RISK OFFICER who stress-tests both for what we'd most regret. Have each argue at full strength — no hedging. Then act as the CHAIR: weigh the arguments, name the single crux this turns on, give your verdict with the reasoning, and tell me what would change it.

You've got it when…

The agent produced genuinely opposed cases — not a both-sides hedge — named the crux the call turns on, and gave a reasoned verdict you could act on. You can see why it landed where it did, and what new fact would flip it.

Quiz — did it land?

Your tutor checks these before marking the lesson complete:

  1. Why does assigning a bull and a bear beat asking one agent for a "balanced" answer?
  2. How is a decision panel different from the multi-model review in the last lesson?
  3. Run a real decision through the panel — what crux did the chair name, and what would change the verdict?