Guide · Agent skill
How to use the Harness Evolve agent skill (evolve)
Updated · written with AI from the project's own documentation and checked against our catalogue data
Harness Evolve is an agent skill that helps an AI agent improve a multi-agent setup, called a harness, that it has already built for your project. It collects feedback on past runs, generalizes it into changes to agents, skills, or the orchestrator, and records each change in a history table. This guide explains what the skill does, how to install it, how to trigger it, and what it needs before you rely on it.
Harness Evolverevfactory/harness · Apache-2.0 Download the skill Quick start
- 1Download the ZIP. Download the evolve skill ZIP from this page.
- 2Unzip it into a skills folder. Unzip into ~/.claude/skills/ for personal use, or into a project's .claude/skills/ for one project.
- 3Or upload it in Claude. In the Claude apps, go to Settings → Capabilities → Skills and upload the ZIP.
- 4Confirm your harness files exist. Check that the project has .claude/agents, .claude/skills and CLAUDE.md, which the skill reads.
- 5Describe the result and the feedback. Tell the agent what went wrong or what you want changed, using a phrase such as 하네스 회고 or 이 피드백 하네스에 반영해줘.
- 6Review the proposed change. Check the generalized change, its stated reason, and the change history entry before you accept it.
What the evolve skill makes the agent do
An agent skill is a folder with a SKILL.md file that tells the agent how to handle a certain kind of task. The evolve skill covers one task: improving an existing harness based on how it has actually performed. The skill describes the harness as something that changes over time. It captures what worked and what did not, feeds that back into the harness, and so the next run starts from a better draft.
The skill works in five phases: collecting the delta, classifying feedback and mapping it to the file that should change, generalizing and applying the change, updating the change history and verifying it, and then reporting the results to you.
- Reads .claude/agents/, .claude/skills/ and CLAUDE.md, including the change history table.
- If the project is a Git repository, reads the history of harness files to see what changed and when.
- If a _workspace/ folder exists, checks whether the outputs from recent runs are where the orchestrator expects them.
- Maps feedback types such as output quality, agent roles, workflow order, team composition, missed triggers, execution mode and cost to the file that should change.
- Records each change with a date, the target, and the reason in the CLAUDE.md change history table.
Install the evolve skill in Claude
Download the evolve ZIP from this page. Then choose one of the two install routes below, depending on whether you want the skill for all your projects or only one.
For a personal install across projects, unzip the file into ~/.claude/skills/. For a single project, unzip it into that project's .claude/skills/ folder. You can also upload the ZIP in the Claude apps under Settings → Capabilities → Skills.
- Personal install: unzip into ~/.claude/skills/
- Project install: unzip into the project's .claude/skills/
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
~/.claude/skills/Trigger the skill with natural requests
You do not need a special command. The skill's description lists phrases that should activate it, so you can ask in plain language after a run that did not go well. The source lists examples such as '하네스 회고' (harness retrospective), '하네스 진화' (harness evolution), '하네스 피드백 반영' (apply harness feedback), '하네스 개선' (harness improvement), '결과가 아쉬웠어 하네스 고쳐줘' (the result was disappointing, fix the harness) and '하네스 레슨 정리' (summarize harness lessons).
If the agent does not pick up the skill, name it directly in your request or check that the folder is in one of the install locations above.
- Say what went wrong and what you want changed, for example: fix the harness because the result was disappointing.
- Ask for a retrospective after a run: 하네스 회고 (harness retrospective).
- Give specific feedback to apply: 이 피드백 하네스에 반영해줘 (apply this feedback to the harness).
Examples of feedback the skill can act on
The skill is most useful when you can point to a concrete result. It asks for general principles rather than one-off patches, so describe the pattern you saw and why it was a problem. The source gives these examples of feedback and the change it maps to.
When a change would add or remove an agent, or redesign the architecture, the skill does not do it itself. It points you to the harness skill instead.
- 'The analysis is too shallow' maps to a skill change that adds a depth standard.
- 'Security review is also needed' maps to an agent change, handled through the harness skill.
- 'Verification should come first' maps to a change in the orchestrator's phase order.
- 'These two agents could be merged' maps to a team composition change in the orchestrator and agents.
- 'This phrasing does not trigger it' maps to an expanded skill description.
- 'Too many tokens are used' maps to a smaller default scale in the orchestrator.
What you need before you start
The skill is meant for an existing harness. It expects a project with .claude/agents, .claude/skills, and CLAUDE.md files. Git is optional, but if your project is a repository, the skill uses Git history to read how the harness changed. The source states that the skill runs no scripts and uses no network access.
The skill is not a way to build a harness from nothing. For that, use the separate harness skill.
- An existing harness with .claude/agents, .claude/skills and CLAUDE.md files
- Git, optional; used to read change history when the project is a repository
- Runs scripts: no
- Uses network: no
Safety notes and practical tips
Because the skill edits the files that define your agents, review its changes before you accept them. The skill is designed to apply one change at a time and check each one, so you can see which change had which effect. If a new change would undo an earlier change in the history, the skill is meant to ask you before it proceeds.
Keep the change history. The source says the history is an asset: it is what makes a future harness in the same domain start from a stronger draft, so do not delete it.
- Give feedback one item at a time so each change can be judged on its own.
- Keep the language of each file consistent. The skill writes changes in the language already used in the target file.
- Review the CLAUDE.md change history after each evolution to confirm it matches the actual files.
- If you change a description, the skill checks triggering with at least three should-trigger and three near-miss examples.
What people use it for
- Tuning a multi-agent content pipeline after a draft came out too shallow or too long.
- Turning repeated fix requests into a rule change in the relevant skill, so the same fix is not needed next time.
- Fixing a trigger that did not fire, by expanding the skill description with the phrasing you actually use.
- Keeping a dated record in CLAUDE.md of why each agent or workflow setting changed.
- Reducing token use by shrinking the default scale of the orchestrator after a run was too expensive.
Frequently asked questions
Is the evolve agent skill free?
The source for this page lists the licence as Apache-2.0 and does not mention any price. Check the repository's licence file for the terms that apply to your use.
How do I install the evolve skill?
Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder. You can also upload the ZIP in the Claude apps under Settings → Capabilities → Skills.
What does the evolve skill require?
It requires an existing harness with .claude/agents, .claude/skills and CLAUDE.md files. Git is optional, and if the project is a repository the skill uses it to read change history.
Does the evolve skill run scripts or use the network?
According to the source, the skill runs no scripts and uses no network access. It works by reading and editing the harness files in your project.
Can evolve build a new harness or add agents?
No. Building a new harness, redesigning its structure, or adding agents is handled by the separate harness skill. Evolve focuses on adjusting an existing harness.
Why does the skill ask for general principles instead of quick edits?
The source says that narrow changes tied to one case lead to overfitting and turn the skill into a pile of rules. The skill generalizes feedback to the underlying principle and records the reason for each change.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use the evolve skill if you built an agent harness in Claude, have used it on real work, and want to tune it while keeping a documented history of why it changed. Pick something else if you are building a new harness or redesigning its architecture, since the separate harness skill covers that, or if you only want a quick one-off edit to some content.
Download the skill