Agent-on-Agent Review
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
A multi-agent quality-control pattern in which one AI instance generates output and a second AI instance critiques it, used to improve the reliability and quality of automated agent workflows.
Overview
Agent-on-agent review, also called the generator-critiquer pattern, involves pairing two AI agents in a feedback loop: one agent produces an output, and a second agent evaluates or critiques that output before it is accepted or acted upon. The pattern is borrowed from established practices and applied within agentic AI systems to provide an automated quality check without requiring human intervention at each step.
Application in Practice
In the context of building proactive agent workflows with Claude Code, this pattern is recommended as a deliberate design investment. The framing — "invest in agent-on-agent review... the generator-critiquer pattern, this is something that we've borrowed here" — suggests the approach is an intentional architectural choice rather than an incidental feature, and that it demonstrably improves output quality in automated routines. (Build a proactive agent workflow with Claude Code, 16:52)
Significance
The pattern addresses a core challenge of autonomous agent systems: without human review at each step, errors or low-quality outputs can propagate downstream. By introducing a dedicated critic agent, the system can self-correct within the loop, making it particularly relevant for long-running or fully automated pipelines where evals and oversight mechanisms are otherwise limited.