Agent Harness Lab

Companion project for the Harness Engineering book.

This is a small support and issue-management app used to show how a weak agent setup can become more reliable through better repository guidance, task briefs, permissions, checks, and review.

This repository starts in a weak but believable state:

• tasks are vague
• repository guidance is light
• permissions are broad
• completion evidence is weak

The later-state example artifacts live under docs/, reviews/, and failures/.

App Surface

• ticket list
• ticket detail view
• status updates
• small JSON API
• focused tests

Install

npm install

Run

npm start

Then open http://localhost:4300.

The root page shows a small dashboard, ticket links, and example API routes. The JSON endpoints are:

• GET /tickets
• GET /tickets/:id
• POST /tickets/:id/status

Test

npm test

For TASK-100, the most relevant check is tests/tickets-status-update.test.js.

Repo Shape

• app/: page behavior and entry points
• api/: API handlers
• lib/: shared ticket logic
• tests/: focused automated checks
• docs/: later-state harness examples
• tasks/: weak starting tasks and stronger brief examples

Reading Path

If you are using this repo alongside the book, start with:

1. tasks/TASK-100.md
2. docs/REPO-GUIDE.example.md
3. tasks/TASK-100-BRIEF.example.md
4. docs/CONTEXT-PACKET.example.md
5. docs/PERMISSION-POLICY.example.md
6. docs/TASK-PLAN.example.md
7. docs/VERIFICATION-CHECKLIST.example.md
8. docs/COMPLETION-EVIDENCE.example.md
9. reviews/REVIEW-CHECKLIST.example.md
10. reviews/APPROVAL-MATRIX.example.md
11. failures/FAILURE-CATEGORIES.example.md
12. failures/FAILURE-LOG.example.md
13. docs/PROGRESSION.example.md

Before And After

The easiest way to use this project is as a comparison exercise.

Before

Start with the weak starting state:

1. read tasks/TASK-100.md
2. open the app at http://localhost:4300
3. inspect the relevant code in app/, api/, and lib/
4. notice what is missing:
   • vague task
   • no strong repo guide
   • no explicit scope boundary
   • no clear completion evidence

After

Then compare that with the later-state harness artifacts:

1. docs/REPO-GUIDE.example.md
2. tasks/TASK-100-BRIEF.example.md
3. docs/CONTEXT-PACKET.example.md
4. docs/PERMISSION-POLICY.example.md
5. docs/TASK-PLAN.example.md
6. docs/VERIFICATION-CHECKLIST.example.md
7. docs/COMPLETION-EVIDENCE.example.md
8. reviews/REVIEW-CHECKLIST.example.md
9. reviews/APPROVAL-MATRIX.example.md
10. failures/FAILURE-CATEGORIES.example.md
11. failures/FAILURE-LOG.example.md
12. docs/PROGRESSION.example.md

The point is not that the repo fully transforms itself. The point is to compare the same small task under a weak harness and a stronger harness.

Practical Use

You can use the project in three simple ways:

1. read the weak task and ask how an agent might drift
2. compare it with the stronger artifacts and see what changed
3. use the repo as a small sandbox for trying your own task briefs, checks, and review rules

License

MIT

Books by the Authors

Scan to check out our books on Amazon