Warp turns human feedback into self-improving skills for its Claude coding agents

The problem: feedback that evaporates

According to a post on Anthropic's Claude blog, surfaced via Hacker News, Warp arrived at its self-improvement loop only after suffering the alternative. The company — an AI-powered terminal and agentic development environment founded in 2020 by Zach Lloyd — found that its internal code review agent produced unhelpful comments and low-quality output. A prompt that handles roughly 80 percent of a recurring task sounds fine until the remaining 20 percent makes the agent noisy and annoying to work with.

The first fixes were manual: engineers rewrote the prompt after observing failures and enriched context files such as AGENTS.md. Both helped; neither scaled. The deeper diagnosis was structural — feedback given to an agent, whatever its purpose, is discarded when the session ends, so every user correction has to be relearned and nothing compounds.

The stakes are not small. The post lists Warp's scale: $73 million raised, around 800,000 monthly developers, use by 56 percent of the Fortune 500, 10 million Claude Code sessions run inside Warp to date (more than 400,000 per week), and 40 million total agent conversations.

Two skills and a human in between

Warp's answer, built on the Claude Platform alongside Rust, Golang, GitHub Actions and its internal orchestration platform Oz, is a loop made of two "skills" — file-based encodings of knowledge that stay out of the raw prompt and are consulted as needed.

The inner, or base, skill holds the functional domain knowledge. When a pull request opens, the code review agent runs against that skill and its context to produce the review. Humans then rate the output — a thumbs up at minimum, but Lloyd argues detailed corrections are far more valuable, for instance explaining that a suggested variable rename violates the codebase's naming convention for globals.

The outer, or improver, skill is an observer agent that runs on a schedule rather than per task. It pulls the accumulated human feedback, compares what the agent suggested against how people actually responded, and proposes a small, focused edit to the base skill.

Because skills are plain files, agents edit them well, and the edits flow through an ordinary pull-request and code-review workflow. Once merged, the next run of the inner skill inherits the improvement. Warp now runs separate spec-writing, review and triage agents across its entire open-source repository, each carrying its own loop.

The triage agent, end to end

The issue triage agent shows the machinery concretely. A GitHub Action fires whenever someone files an issue; the agent assesses complexity and feasibility, assigns labels and suggests a fix direction, guided by an inner skill defining what each label means and how to research the codebase before acting.

In the example in the post, the agent did solid work but missed one label, "ready to spec," which signals that contributors can start building product and technical specs against the issue. A maintainer left feedback directly on the issue — where the work was already happening — spelling out both what was expected and why.

The improver skill then ran inside Oz as a scheduled "update triage" job. It authenticated to GitHub, executed a Python script bundled with the skill to fetch recent feedback-bearing issues, summarized them into a JSON file, and read that back into context. From there it isolated the maintainer's concrete signals and opened a PR containing the smallest edit to the skill that captured them.

Guidance for writing improvable skills

Lloyd's team offers rules of thumb. Write principles rather than exhaustive rules: directing an agent as if instructing "a smart person" generalizes better than enumerating variable-naming conventions. Explain the rationale behind each instruction so the agent can reason instead of pattern-matching. Make feedback nearly effortless to give, captured in PR or issue comments with no separate submission step — in his words, "low friction is what keeps signal flowing." Keep skills small, referencing scripts and resource files rather than stuffing everything into context. Prefer feedback quality over volume, though Warp's thousands of reviews across hundreds of contributors suggest both help. And invest in the improver skill itself: apart from the domain-specific portion, it is largely reusable from one agent to another.

Why it matters

Most agent deployments today are stateless learners: whatever a user teaches the agent in a session disappears when it closes, so quality plateaus at the level of the original prompt. Warp's pattern converts that ephemeral feedback into versioned, human-reviewable artifacts, letting an agent improve the way software does — through diffs, PRs and review. Notably, it needs no fine-tuning and no bespoke memory infrastructure, just files, a scheduler and a standard review workflow on top of ordinary model access. The limitation is symmetric: the loop only compounds as fast as humans explain themselves, since a binary thumbs up never says why.