TL;DR
New pattern for closed-loop agent workflows that review, repair, and validate artifacts with structured feedback across multiple iterations.
Key Points
- Three-phase architecture: Review identifies issues, Repair applies focused edits, Validate runs checks and produces feedback for next iteration
- Structured handoffs between phases enable auditable maintenance workflows with machine-readable outputs (JSON schemas, record.json files)
- Practical beyond documentation: applicable to code modernization, protocol optimization, regulatory remediation, and knowledge base updates
- Multi-iteration convergence: demonstrated 3-pass flow where simple cases clear in iteration 1, complex cases in iteration 3 with decreasing validation deltas
Why It Matters
This pattern solves a critical problem in agentic systems: how to make AI-driven maintenance auditable and trustworthy. By separating judgment (review) from proof (validation) and requiring structured evidence at each step, teams can deploy agents to fix documentation, update code, and maintain compliance without blind trust in single-pass outputs. The iterative approach with explicit stop conditions makes agentic workflows practical for production systems.
Source: developers.openai.com