Warp publishes the shape of a feedback-to-skill loop, without measuring it

Warp's internal code review agent was producing comments engineers found unhelpful. Rewriting the prompt by hand worked but did not scale, and improving context files helped without fixing it. The architecture they landed on has two skills with a human between them. An inner skill holds the domain knowledge and does the work. An outer improver skill runs on a schedule, not per task: it pulls accumulated human feedback, compares what the agent suggested against how humans responded, and proposes one small edit to the inner skill. Because skills are plain files, that edit arrives as a pull request a human reviews, approves and merges.

The worked example is their issue triage agent, published at warpdotdev/warp-agents-demo-github-issue-triage. The triage agent missed a "ready to spec" label; a maintainer left feedback on the issue saying both what he expected and why; the improver pulled recent issues carrying feedback via a bundled Python script, summarized them to JSON, and opened a pull request teaching the inner skill that rule. Founder Zach Lloyd's guidance for writing these: "Construct the skill as though you're instructing a smart person, not like you're programming a computer," and on feedback capture, "Low friction is what keeps signal flowing."

The piece reports no measurement. Warp's first-party figures are about the company, not the loop: 800K monthly developers, 10M Claude Code sessions run inside Warp to date, more than 400K per week. There is no before and after on review quality, so the architecture is documented and the improvement is asserted.

Why it matters: Warp says the improver skill is largely the same across agents once you strip out the domain knowledge. The human gate is equally important: the loop edits a file and files a pull request, so a person approves every change to the agent's own instructions. Warp's guidance assumes some feedback will be wrong: "Assume it will be." It also says to build the harness first, then let the agent tune against it. Skip either and you have built a mechanism for writing one reviewer's preference into a permanent rule for everyone.