Developer builds self-improving code pipeline using three collaborative AI agents

A developer has created a multi-agent AI pipeline in which one AI agent writes Python code, a second scores it, and a third refines it based on structured feedback — all running in an automated loop. The system uses Anthropic's Claude models, with the generator and refiner powered by claude-opus-4-8 and the scorer using the faster, cheaper claude-haiku model. Code is accepted only when it scores 9.6 or above out of 10, with a maximum of three refinement attempts before the pipeline exits with an error. A key design insight was passing the full history of previous attempts to the scorer to prevent it from penalising changes it had already rewarded in earlier rounds. Once the code clears the threshold, it is written to a temporary file and executed as a child subprocess, completing the automated cycle.