Warp builds self-improving agents on Claude

Warp, the AI-powered terminal and agentic development environment, has developed a simple yet effective method for creating self-improving agents. This approach, centered around an Agent Skills-based framework, allows feedback to compound over time, refining and enhancing agent output continually. The key to this innovative technique lies in skills, file-based encodings of knowledge that keep instructions separate from the raw prompt.

The process begins with a first-pass prompt, which must reliably handle recurring tasks. However, even a prompt achieving 80% accuracy can lead to a poor user experience. Warp learned this lesson the hard way, particularly with their internal code review agent. Engineers complained about unhelpful comments and low-quality output, indicating a need for improvement.

The team initially tried stopgap solutions, such as manually rewriting prompts based on observed failures and improving context files. However, these methods were far from complete fixes. Realizing that feedback typically disappeared when the session ended, they devised a solution: an Agent Skills-based framework to create self-improving agents.

The central technique involves a self-improvement loop using skills, which are file-based encodings of knowledge. This keeps instructions out of the raw prompt, allowing Warp agents to evolve a self-improving agent architecture consisting of two skills: an inner/base skill holding functional domain knowledge and instructions, and an outer/improver skill observing accumulated human feedback and proposing small edits to the base skill.

Human feedback is a critical component of this loop. For example, in code review, a thumbs up could affirm a good comment, while detailed reasons for dissatisfaction could guide the agent on how to improve. The outer/improver skill, running on a schedule, pulls this accumulated feedback, compares the agent's suggestions with human responses, and proposes edits to the base skill.

These updates, residing in plain files, can be easily reviewed, approved, and merged into the next run of the inner skill, thus continuously refining the agent's output.

Warp has implemented this pattern across its entire open-source repository, with separate agents for spec-writing, review, and triage, each carrying their own self-improvement loop. The issue triage agent, for instance, automatically analyzes new GitHub issues, assigns labels, and suggests fix directions based on an inner skill file containing domain knowledge.

When a maintainer notices a gap in the agent's performance, they leave actionable feedback, which the outer improver skill then refines and updates the base skill accordingly.

This self-improving agent skills framework not only streamlines the development process but also empowers humans to maintain control over what changes are made. By leveraging the power of file-based skills, Warp has built a robust system for creating self-improving agents that can continually enhance their performance over time.