How a Worker-Evaluator Loop in LangGraph Lets AI Agents Check Their Own Work

A developer built an AI agent system called Wingman that uses two separate roles — a worker that completes tasks and an evaluator that judges whether the output meets defined success criteria. The evaluator uses structured output, returning typed boolean flags instead of free-text responses, keeping routing logic clean and reliable. If the criteria are not met, the worker retries with specific feedback injected into its system prompt, rather than attempting the task blindly again. A hard cap of five turns prevents runaway retries, with the loop exiting early if criteria are met or the agent determines it needs user input. The architecture was designed for LangGraph and later tested on AWS Lambda behind API Gateway, where specific deployment issues emerged.