Critic Pattern

What is the critic pattern?

A writer rarely spots their own typos, but a fresh editor catches them instantly. The critic pattern is that idea, applied to agents: one agent produces the work, and another exists purely to poke holes in it — pointing out problems and suggesting fixes.

How is it different from self-reflection?

Self-reflection is checking your own work
When the same model re-reads its own output, it tends to grade itself generously and miss obvious errors.
A critic is a second set of eyes
An independent reviewer isn't locked into the momentum of the production step, so it catches logic holes, factual mistakes and formatting slips more easily.

How to set it up

Separate the roles
The generator: build the thing. The critic: only find faults, don't fix them.
Give the critic a checklist
Something like "does it meet the brief, did it invent facts, is the tone right" — and have it tick each item.
Iterate
Generate, criticize, revise, criticize again — until the critic finds nothing major, or you hit your iteration cap.

What it's worth — and what it costs

The critic pattern measurably lifts quality on complex work — code, contracts, copy — where mistakes are expensive. The trade-off: extra rounds of calls, more time and more tokens.

Bottom line: the critic pattern puts one AI to work and another AI to sign off.