AI Coding Agents Self-Improve with Accumulated Behavioral Rules.

This research presents a novel closed-loop framework designed to enable AI coding agents to continuously improve their performance by learning from human feedback. The core idea is to transform every accepted human review comment into a persistent behavioral rule, which is then added to an accumulating rule set. This rule set, stored in a version-controlled instruction file, allows the agent to self-detect and correct error classes it previously made. The framework integrates a self-review checklist, executed by the agent before code submission, and automated validation to ensure the integrity of the growing rule set. In a deployment across a microservices platform, the rule set expanded significantly, incorporating numerous behavioral rules, language-specific standards, and a self-review checklist derived from real-world feedback. Empirical results from 11 working sessions demonstrated that these accumulated rules effectively shift human review effort from low-level correctness checks to higher-level design validation. Crucially, the framework achieved a measured 0% recurrence rate for error classes covered by the rules, and these learned behaviors successfully transferred across different agent interfaces. This approach offers persistent, cross-session learning without requiring model weight updates, addressing a critical gap in existing LLM learning methods.