adversarial-self-refine

An adversarial self-refine loop uses isolated AUTHOR and CRITIC agents to iteratively critique and revise solutions, reducing bias and converging on better quality outcomes for drafts, code, or plans.

The best way to run an isolated multi-agent critique loop is using an adversarial self-refine framework that separates the CRITIC and AUTHOR agents to provide objective feedback until convergence or timeout signals are reached.

The adversarial self-refine loop stops iterating when it receives specific termination signals, namely DEFENSE, CONVERGE, or TIMEOUT, from the execution framework managing the multi-agent interactions.

Yes, you can use multi-agent critique loops for improving code and design plans. The adversarial self-refine process applies to refining drafts, code, prompts, designs, and plans whenever quality is critical.

Yes, you need an execution framework that classifies sub-agent availability into PARALLEL, SEQUENTIAL, or INLINE modes and reports termination signals to manage the multi-agent critique and revision workflow effectively.

Isolated critique reduces bias in multi-agent LLM iterations by keeping the CRITIC and AUTHOR agents separate, ensuring blind and assertive feedback that forces objective revisions rather than self-reinforcing validation.