Warp builds self-improving agents on Claude

TL;DR Highlight

Splitting a base skill and an improver skill let Warp's agents improve from accumulated feedback

Who Should Read

A backend developer whose code review or issue triage agent leaves noisy comments and who needs a way to feed back human corrections

Core Mechanics

Feedback vanishes when the session ends
- Warp's internal code review agent left unhelpful comments and engineers complained.
- Manually rewriting the prompt and improving AGENTS.md helped but did not fully fix it.
- The root cause was that feedback given to an agent disappears when the session ends.
Two skills form the loop
- The base (inner) skill holds domain knowledge and instructions.
- The improver (outer) skill runs on a schedule, gathers human feedback, and proposes a small edit to the base skill.
- Skills are file-based knowledge kept out of the raw prompt.
Improvements ship through PRs
- Skills are plain files, so agents are good at editing them.
- Edits go through normal PR review, approval, and merge, and the next run inherits them.
- The final human approval closes the loop and keeps a person in control.
Issue triage agent example
- A GitHub Action fires an agent on each new issue to assess complexity and feasibility, assign labels, and suggest a fix direction.

Community Reactions

Opinion that self-modifying loops need more risk controls
- One commenter said letting an agent edit its own instructions crosses a risk threshold the article underplays.
- They noted PR approval is framed as checking if the idea is good, not as defense against prompt injection.
Pushback that it offers no determinism
- One view held that human-in-the-loop plus principles-not-rules gives no guarantees for deterministic workflows.
- Another quipped that skills are just markdown files the agent can ignore at any time.
Experience building similar loops locally
- One commenter updates a review skill when human reviewers make valid criticisms the agent missed.
- Another uses CLAUDE.md to log every decision and has the agent challenge them with it.
Slowdown with more rules and model-upgrade concerns
- One commenter said the agent seemed slower after adding principles and wondered about a performance and cost tradeoff.
- Another noted that upgrading models can make some rules redundant and force rework.

How to Apply

Run an improver prompt on your review agent
- Today, paste last week's PR review comments and human reactions into the prompt below.
- Include the current review skill or AGENTS.md and ask for one smallest edit.
Make PR comments the feedback channel
- Tell the team to leave feedback with reasons in PR or issue comments, not a separate form.
- Put a script in the skill folder that collects only feedback that includes reasons.
Schedule the improver and ship only via PR
- Run it periodically with something like GitHub Actions to open PRs editing skill files.
- Disable auto-merge and require human approval.
Rewrite skills around principles and reasons
- Replace long naming-rule lists with principles like look for repeated code.
- Add a one-line reason next to each principle.

Try It

copy & run

# Paste into Claude or ChatGPT as-is (simplified version of Warp's improver skill pattern)

You are the improver for a code review agent. Look at the current skill and recent review feedback below and propose one smallest edit to the skill.

[Current skill]
{{current review skill or AGENTS.md content}}

[Recent agent comments and human reactions]
{{comments, human reactions, reasons}}

Rules:
1. Write principles, not lists of rules, and add a one-line reason.
2. If the feedback appears only once or is weakly justified, do not edit and explain why.
3. Output in this order: (a) feedback signals used (b) the edit as a diff (c) a draft PR description.

Next step: apply the edit to the real skill file, rerun the review on the same PR, and compare comment quality.

Terminology

skillA guidance file the agent looks up when needed. Like keeping a manual in a drawer instead of reciting it every time.

improver skillA skill that edits other skills. It runs on a schedule as an observer, reads accumulated feedback, and proposes edits.

progressive disclosureLoading resource files only when needed instead of putting everything in context upfront. Like reading the table of contents and opening only the chapter you need.

OzWarp's agent orchestration platform, where the scheduled improver agent runs.

AGENTS.mdA context file that tells agents about a project's background and rules.