@tank/adversarial-agent-coach

To verify AI agent answers, apply adversarial review to detect unsupported claims, demand source-backed evidence, and run verification loops. This contradiction check process rewrites weak responses with calibrated uncertainty for higher agent reliability.

Reflexion is a retry and revision loop method where failing agent workflows are iteratively improved. It uses structured claim scrutiny and repeatable validation across revisions to tighten reasoning and ensure defensible outputs.

Set up multi-agent critique by splitting prompt tuning into builder, skeptic, and verifier roles. This adversarial review framework forces evidence demands and contradiction checks, yielding eval-first improvements based on validation rather than tone.

Yes, you can use ReAct retries to fix suspicious agent responses. The workflow identifies the weakest claims, applies contradiction checks, adds caveats, and rewrites the output with source-backed support and verification steps.

Yes, adversarial review works without external dependencies. It applies contradiction checks, evidence demands, and verification loops entirely through prompt-driven logic to strengthen weak AI agent answers.

Avoid adversarial review for low-stakes tasks where response speed outweighs reliability. The contradiction checks and verification loops add latency, making it unsuitable when calibrated uncertainty and source-backed support are not required.