Agentic Research

Debate / adversarial

Several positions attack each other; disagreement surfaces the weak points.

Shape:Opposing agents, a third adjudicates

When to use it

High-stakes conclusions needing stress tests: investment decisions, architecture choices, risk assessment. The value is exposing weak arguments, not producing consensus.

When not to

Questions with objective answers. Debate does not make arithmetic right; it makes a wrong answer sound more convincing.

How it fails

  • The most expensive pattern, with benefits that are hard to quantify.
  • Converging on consensus rather than correctness. Several same-origin models reinforce one another's error.
  • The assigned opposition does not try hard, because it knows it is playing a role.

Design notes

Give the opposition a different information surface from the proposer — different retrieval results, different assumptions — or it is one model talking to itself. The adjudicator must not also be a participant.

Origin

未核實到可引用的首次公開實作日期;不填。

Related articles on this site