Self-improving coding agents cut bug recurrence to 0% with persistent rules

A new research paper from Aditya Aggarwal and Nahid Farhady Ghalaty introduces a closed-loop framework that lets AI coding agents self-improve by accumulating behavioral rules from human review feedback. Instead of fine-tuning model weights—which is costly and can cause catastrophic forgetting—the system codifies each accepted review comment as a persistent rule stored in a version-controlled instruction file. The agent runs a self-review checklist before every code submission, catching errors it has learned to avoid. Automated validation ensures the growing rule set remains consistent and actionable.

Deployed on a production microservices platform with over 35 services, the rule set expanded from 5 to 18 behavioral rules, along with 15+ language-specific standards and a 15-item self-review checklist—all derived from real code reviews. In 11 recorded sessions covering generation, PR review, incident investigation, and refactoring, the framework achieved a 0% recurrence rate for ruled-against error classes. Review effort shifted from low-level correctness checks to higher-level design validation. The approach outperforms prior methods like Reflexion, ExpeL, and Voyager by enabling persistent cross-session learning without weight updates, and works on production codebases rather than synthetic benchmarks.