Warp Builds Self-Improvement Loop for Agents Based on Claude

AI terminal Warp has unveiled a self-improvement loop architecture that continuously enhances agent performance by leveraging the Claude platform. Initially, the team relied on manual prompt adjustments or context file improvements, but recognized limitations in scalability due to feedback being lost at the end of sessions, and addressed this issue.

The core of the approach is a two-stage structure utilizing file-based skills.

• Internal/Basic Skills: Contain domain knowledge and guidelines, referenced when performing actual tasks (e.g., code reviews).
• External/Improvement Skills: Observer agents executed on a schedule that analyze accumulated human feedback and propose modifications to the basic skills.

This method integrates naturally with existing development processes because skills are in standard file formats, allowing for review, approval, and merging via PR workflows. Warp is applying this pattern across specification writing, review, and triage agents in its open-source repositories.

The Warp team emphasizes principle-based instructions, explanations of the "why", and low friction for providing feedback to create effective skills. In particular, feedback is designed to be automatically collected from existing workflows such as PR comments, ensuring the continuity of signals.