Warp keeps agent feedback from vanishing by writing it back to a skill file

The everyday failure of production agents is that a good correction dies with the session. Someone tells the agent it missed something, the agent fixes it that once, and the next run starts from zero. Warp, a terminal and agentic development tool that says it runs 400,000 Claude Code sessions a week, describes a pattern to stop that leak. It uses two skills. A base skill holds the domain knowledge and instructions an agent works from, for example a code review agent. An improver skill is a separate observer that runs on a schedule, collects the human feedback, compares what the agent suggested against what people actually did, and proposes focused edits to the base skill.

The concrete example is their issue triage agent, which labels new GitHub issues and suggests where a fix might go. When a maintainer pointed out it had missed the "ready to spec" label, the improver skill opened a pull request updating the base skill. That change is reviewed, approved, and merged like any other code, and only then does the next agent run inherit it. A person stays in the loop before any improvement takes effect. Warp's advice for writing these skills is to state principles rather than rules, as founder Zach Lloyd puts it, instruct a smart person instead of programming a computer, explain the reasoning behind guidance so it generalizes, capture feedback where people already work such as PR comments, and keep each skill small.

Why it matters

If you run agents in production, this is a copyable recipe for the part that usually breaks: making corrections stick. Because the improvement is a diff to a skill file and goes through normal code review, you get persistence and an audit trail instead of prompt edits nobody can trace.