Warp Automates Prompt Engineering With Self-Improving AI Agent Framework

Warp has implemented a self-improvement loop for its AI agents using the Claude Platform to eliminate the need for manual prompt engineering. The system allows agents specializing in issue triage, spec-writing, and code review to evolve based on real-world developer feedback.

To achieve this, Warp utilizes a two-tier "skills" architecture. An "inner" or base skill contains the specific domain knowledge required for a task, while an "outer" improver skill acts as an observer. This improver agent analyzes human feedback to propose edits to the base instructions. According to Warp CEO Zach Lloyd, file-based skills allow agents to look up encoded knowledge during a task rather than having that information crammed directly into the prompt.

The Feedback Loop

The system was developed to solve the problem of "statelessness," where AI agents lose the context of feedback once a session ends. Warp initially encountered this issue with internal code review agents that were producing unhelpful or low-quality comments. To fix this, Warp moved away from manual prompt rewriting and toward a systematic knowledge management system using their orchestration platform, Oz, and Claude's skills framework.

Integration into the developer workflow is central to the process. When the improver agent identifies a necessary refinement, it opens a pull request (PR) to update the skill file. This ensures that while the AI proposes the improvement, a human developer must review and merge the change before it goes live. Lloyd noted that maintaining "low friction" is essential to keep the signal flowing from users to the system.

Industry Implications

This approach provides a scalable model for "agentic loops," where AI systems compound knowledge over time. By treating agent instructions as code—subject to version control and PRs—companies can automate the tedious process of prompt tuning while maintaining strict human oversight. This transforms prompt engineering from a manual art into a systematic software engineering process.

Scale and Adoption

The framework supports a significant user base, as Warp reports 800,000 monthly developers and notes that 56% of the Fortune 500 use its terminal. The scale of AI integration is further evidenced by the 10 million Claude Code sessions that have run inside Warp to date, with current volume exceeding 400,000 sessions per week. Additionally, the company has seen 40 million total Warp Agent conversations.

Moving forward, the industry will be watching whether this file-based skill architecture can be applied to more complex, multi-step autonomous workflows beyond the terminal environment, and how the ratio of AI-proposed versus human-authored instructions shifts as the improver agents mature.