Turn accepted review comments into persistent behavioral rules
Closed-loop framework that codifies review feedback into versioned rules and a self-review checklist—0% recurrence for ruled-against errors in deployment.
Source: Self-Improving AI Coding Agents Through Accumulated Behavioral Rules: A Closed-Loop Framework
Coding agents repeat the same mistake classes because corrections die in chat. This paper’s fix is boring and effective: every accepted review comment becomes a persistent behavioral rule in a version-controlled instruction file, plus a pre-submit self-review checklist and automated integrity checks as the rule set grows.
In a 35+ service microservices deployment, rules grew from 5 to 18 (plus language standards and a 15-item checklist). Across 11 recorded sessions they report 0% recurrence for ruled-against error classes and a shift of human review toward design-level concerns.
Why it matters for code review
This is the operational version of Reflexion: review is the teacher, the rule file is the memory. If your org has an agent writing PRs, the minimum viable learning loop is “accepted comment → durable rule → self-check before next PR.”