How to use the Harness Engineering agent skill

What is an agent skill, and what does this one do?

An agent skill is a folder containing a SKILL.md file with instructions that an AI agent reads when a task matches the skill's description. This skill's description says it should be used when designing autonomous agent harnesses, covering research loops, evaluation scaffolds, locked and editable surfaces, durable logs, novelty gates, pruning, rollback, PR preparation, and human approval boundaries.

The skill is a design guide for the system around an agent, not a tool that runs code itself. It helps you decide four kinds of surface: locked (such as evaluation rubrics and merge rules, which the agent may read but not change to score itself), editable (the draft, experiment file, or config under test), append-only (results logs and rejected ideas, which can be added to but not rewritten), and human-controlled (merges, production deploys, credentials, and destructive operations).

Install the harness-engineering skill

Download the ZIP for the skill from this page and unzip it. The skill is one file, SKILL.md. For a personal install across projects, place the unzipped folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder.

In the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project
• Claude apps: Settings → Capabilities → Skills, then upload

Trigger the skill with natural requests

You do not need a special command. The skill is meant to be used when a request matches its description, so describe the design problem you are working on. The skill's own activation list includes building autonomous research or experimentation loops, designing an agent environment with locked metrics and editable code or content, creating PR-producing or background agents, and checking whether an agent can run without frequent prompts.

Name the constraints you care about, such as what must stay locked or where progress should be recorded. The more concrete the request, the more useful the playbook will be.

• Help me design an overnight loop where an agent tunes my config, but it must not change the evaluation script.
• I want a background agent that prepares PRs for review. Which files should be locked and what needs human approval?
• Set up a durable log so a future agent can resume this research run without the chat history.
• Review my agent workflow: can it run for days without a person checking in, and where could it game its own score?

Work through a harness design

The skill's checklist gives a clear order. Start by defining the objective in one sentence. Then mark the locked, editable, append-only, and human-controlled surfaces. Choose the feedback mechanism, which can be a scalar metric, a rubric, deterministic tests, human review, or a combination. Define keep, discard, crash, timeout, and review states, and create a durable thread log before the loop starts.

Before you trust the loop, validate the harness on one known good artifact and one known bad artifact. A harness that cannot tell them apart is not ready for autonomy.

• Define the objective in one sentence.
• Identify locked, editable, append-only, and human-controlled surfaces.
• Choose the feedback mechanism.
• Define keep, discard, crash, timeout, and review states.
• Create a durable thread log before the loop starts.
• Add source refresh, novelty, and pruning rules for long runs.
• Define what the agent may do without asking and what needs approval.
• Validate on one known good and one known bad artifact.

Set up the file layout for a research run

The skill suggests a layout that keeps state in files so any future agent can pick up the work. THREAD.md holds the running log, sources hold the queue and evaluation notes, and logs holds a results file and a rejected-ideas file.

The skill advises TSV or JSONL for append-only machine-readable logs, and Markdown for handoffs and reviewer-facing summaries.

Apply the autoresearch-style loop safely

For optimizing an artifact against a stable evaluator, the skill describes this loop. The agent reads locked context, picks a hypothesis, edits the allowed surface, checkpoints, runs the evaluator, logs the result, keeps the change if it is better, and rolls back if it is worse. The evaluator must sit outside the editable surface, and the agent needs a policy for crashes and timeouts.

The skill also lists guards against metric gaming, such as lock rubrics per run, report per-dimension scores, and route governance changes to human review.

• Evaluator lives outside the editable surface.
• Feedback cadence is fixed enough to compare attempts.
• Failed attempts leave an audit trail.
• Rollback is cheap.
• Agent has a policy for crashes and timeouts.

Know what the skill does not cover

The skill tells the agent to hand off some topics to other skills. General quality gates, regression suites, and outcome metrics without autonomous control surfaces belong to the evaluation skill. Tool schemas and error formats belong to tool-design. Project-level task fit and cost planning belong to project-development. Remote sandboxes and hosted session infrastructure belong to hosted-agents.

Simple one-off chats or short coding tasks do not need this structure. The extra control layer adds overhead without benefit there.

Safety notes: scripts, network and approvals

According to the skill's listing, it runs no scripts and uses no network access, and it lists no requirements. It is only instructions that shape how an agent designs a system.

The skill is explicit about governance. Autonomous agents may draft changes, run checks, and write PR summaries, but they should not merge, deploy, or push without human approval unless the user has explicitly granted permission for that specific action. Unsupervised production deployment without human approval is something the skill warns against. The license is MIT.

• Runs scripts: no
• Uses network: no
• Requirements listed: none
• Agents should not merge, deploy or push without explicit human approval.