Verification Loops & Self-Correction

The difference between an amateur agent and a production system is self-correction. Instead of assuming the first generated output is correct, top agent architectures run an explicit validation check before delivering results to the user.

The Critic / Verifier Pattern

• Unit Test Verification: If the agent generates code, run an automated test runner against the code. Feed compiler or test errors back into the agent to trigger a fix cycle.
• Deterministic Schema Validation: Pass the output through a schema validator like Zod. If validation fails, return the exact schema error to the LLM with instructions to repair the JSON.
• Secondary Verifier Pass: Use a secondary, lightweight model call acting as an objective critic to verify that all constraints in the initial user prompt were satisfied.

const result = OutputSchema.safeParse(agentResponse);
if (!result.success) {
  // Feed exact issue back to agent for deterministic repair
  return agent.repair({
    errors: result.error.issues,
    previousOutput: agentResponse,
  });
}