Back to blogs

August 4, 2025 · 4 min read

Why I Make My Agents Fight

Agreement is cheap. Useful output comes from critique. Making agents argue under a rubric is a practical way to surface blind spots before they become mistakes.

AI WorkflowsEvaluationDecision MakingExecution

Headline Signal

Critique creates reliability

Polite Agents Create Polite Mistakes

If every output sounds confident, you lose the signal that something is wrong.

Critique is what turns a draft into a tool you can trust.

That is why adversarial review is useful: it forces assumptions to show themselves.

This is not about drama. It is about safety and quality.

How to Create Productive Conflict

Give one agent the job of proposing and another the job of breaking.

Provide both agents the same context so the conflict is about reasoning, not missing info.

Define a rubric before they start.

Require the breaker to propose fixes, not just criticism.

Where This Helps Immediately

Use it on plans that touch customers, money, or brand.

Use it on workflows before you automate them.

Use it on hiring decisions where bias can sneak in.

When the output survives critique, you can ship with more confidence and fewer surprises.

  • Define the decision and the constraints.
  • Run a proposer and a breaker.
  • Score against a rubric.
  • Update the plan and rerun.
  • Log assumptions and revisit later.

Rubrics Turn Critique Into an Asset

Over time you build a scenario pack: inputs that represent your real world.

You build a rubric that matches your standards.

And you build guardrails that prevent repeat mistakes.

That is what makes agents useful: not charm, but reliability.

Bottom Line

Use adversarial review on one real decision this week. Define a rubric, force critique, and ship the version that survives the break tests.