Corpus Agentis
The field book to agent ecosystems
The field book to agent ecosystems
Design · Topologies

Adversarial Review

Producer emits, critic rebuts, arbiter (optional) resolves. Loops until convergence or timeout. Widely used in code-review + reasoning-verification.

Verified against arxiv.org

Adversarial Review agent topology diagram

An adversarial-review topology pairs a producer that emits work with a critic that tries to refute it, and optionally an arbiter that resolves the two, looping until the critic can no longer find a flaw. Building the check into the loop raises quality: nothing ships until it survives a dedicated adversary. It is the shape behind self-checking coding and writing agents. In anatomy terms it is two opposed cores plus an optional judge, iterating.

In the anatomy

A producer core emits; a critic core attempts to refute; an optional arbiter core resolves; the loop repeats until convergence; effectors commit only the arbiter-approved output.

At a glance
PatternAdversarial Review
State modelsession
ExampleAnthropic red-team pairs · code-review agents
Common use cases
  • ·Self-checking code and writing agents: generate, then critique.
  • ·Fact-checking and verification loops.
  • ·Any output that must survive scrutiny before it ships.
Strengths
  • ·Higher quality: output must survive an adversary.
  • ·Catches errors a single pass would ship.
  • ·Self-documenting via the critique trail.
  • ·Tunable rigor via the loop budget.
Weaknesses
  • ·More expensive: at least twice the calls per round.
  • ·Can loop without converging.
  • ·A weak critic gives false assurance.
  • ·Latency of the review rounds.
Basis

Description, use cases and trade-offs: Field Book analysis, synthesised from established agent-architecture literature.

Verified against
https://arxiv.org/abs/2502.02533

The corpus does not rewrite vendor documentation. Where copy is quoted verbatim it is marked as such and attributed to its source.