Warp's Agents Rewrite Their Own Instructions — With a Human Still Signing Off

Warp, the developer-tools company running more than 10 million Claude Code sessions and 40 million total Warp Agent conversations through its product, builds agents that improve themselves using a two-layer skill architecture on Claude’s Skills API. An inner “base” skill holds the functional domain knowledge an agent needs for its primary task; an outer “improver” skill is a separate observer agent that runs on a schedule, analyzes human feedback on the base agent’s output, and proposes targeted edits back to the base skill. Anthropic profiled the approach on its blog. Warp — founded in 2020 by CEO Zach Lloyd, having raised $73M, used by 56% of the Fortune 500 with 800,000 monthly developers — routes every proposed skill update through its normal code-review workflow, requiring human approval before a change merges. The agent doesn’t get to quietly rewrite its own instructions; a person still has to sign off on how it evolves.

Warp applies the same pattern beyond a single agent, running separate spec-writing, review, and issue-triage agents across its open-source repository — the issue-triage agent, for instance, analyzes incoming GitHub issues for complexity, assigns labels, and suggests fixes. Anthropic distilled five practices from the experience: write skills as principles rather than rigid rules, explain the reasoning behind instructions rather than just stating them, capture feedback inside existing workflows instead of separate channels, prioritize feedback quality over volume, and invest real design effort specifically in the improver skill.

The timing lines up with the Skills API’s move to general availability alongside Computer Use and the Files API on August 20th, where an Asteroid-built claims-processing workflow using the same APIs cut per-case handling time from 32 minutes to 13 — roughly a 30% cost reduction. Warp’s self-maintaining skill files are a preview of what production teams build once agent instructions stop being a one-time prompt and start needing their own review loop.