How Warp Builds Self-Improving Agents on Claude Using File-Based Skills

Warp, an AI-powered terminal and agentic development environment, developed a pattern for self-improving agents using Agent Skills on the Claude Platform. The core problem was that user feedback typically disappears after a session, preventing agents from learning from mistakes. Warp’s solution involves a loop where an inner/base skill executes tasks, and an outer/improver skill periodically analyzes human feedback to propose edits to the base skill file.

The Self-Improvement Loop

The architecture relies on two distinct skills:

• Inner/Base Skill: Contains functional domain knowledge and instructions. For example, a code review agent uses this skill to generate output.
• Outer/Improver Skill: An observer agent that runs on a schedule. It pulls accumulated human feedback, compares agent suggestions against responses, and proposes small, focused edits to the base skill.

Because skills are plain files, these updates can flow through standard PR/code-review workflows. Once merged, the next run of the inner skill inherits the improvement. Warp currently runs this pattern across its open-source repo, managing spec-writing, review, and triage agents.

Best Practices for Skill Design

The Warp team recommends specific strategies for writing effective self-improving skills:

• Write principles, not rules: Instruct the agent like a smart person rather than programming a computer. For example, "Look for repeated code" is better than exhaustive naming conventions.
• Explain the why: Providing rationale allows the agent to reason and generalize better than rigid instructions.
• Make feedback effortless: Capture feedback where people already work, such as PR comments, with no extra submission steps.
• Prioritize quality over volume: Detailed feedback from senior engineers is more valuable than binary thumbs up/down, though a larger corpus of quality signal helps.
• Keep skills small: Use progressive disclosure by referencing resource files instead of dumping everything into context.

Implementation Example: Issue Triage

Warp’s issue triage agent demonstrates this framework. When a GitHub issue is filed, a GitHub Action triggers an agent to analyze complexity and assign labels. If a maintainer notices a missing label (e.g., "ready to spec"), they leave feedback on the issue. The outer improver skill, running in Warp’s Oz orchestration platform, pulls this feedback, identifies the signal, and opens a PR to update the inner skill file. This ensures the agent correctly applies the label in future instances, keeping a human in the loop for final approval.