Debate and judging

Debate is expensive and occasionally worth it. Its real value is not a better answer, it is a visible argument: the losing side's case is the thing a human most needs to read before making a real decision.

How it works

State the question as a proposition with two defensible sides.
Agent A argues for, agent B argues against, each with the same evidence.
A judge with written criteria picks the stronger case and says why.
Both arguments are kept, not just the winner.
A human reads the losing argument, because that is where the risk they had not considered lives.

Optional: load a real model

See it work

When it pays, and when it does not

Use it when	Skip it when
Real tradeoffs exist and experts disagree	There is a correct answer and you just need it
A human will read the arguments	Nobody will read anything but the verdict, which makes the cost pure waste
The decision is worth several model calls	The task runs thousands of times a day
You want the risks named explicitly	You need speed more than completeness

How it fails

Confident nonsense on both sides

Two models can argue fluently for two wrong positions. Ground both in the same evidence set or you get theatre.

Judge bias to length

Judges reward longer arguments. Cap length and score against explicit criteria.

Debate as decoration

If nobody reads the losing argument, you paid three times for one answer.

What it costs

Tokens	Three to five calls minimum, and long ones. The most expensive pattern in this track.
Latency	Slowest pattern here unless the two sides run in parallel.
Value	Real but narrow. Most teams should use a critic instead of a debate.

Keep the losing argument. It is the part that changes decisions, and the pattern is a waste of money without a human who reads it.

Related: Multiple agents · Reflection · Sampling and voting · all patterns · agent jobs

Free from AI School - no signup, everything runs in your browser.