Skills and hooks that make Claude Code finish what it starts

Description and use

The agent-harness scripts and hooks ensure agents only finish tasks after evidence verification and require owner approval for sensitive actions.

Problem

Agents often report completion without verifying that all required steps, such as URL checks, request completions, or automated next steps, have actually been performed.

Outcome

Agents now only declare completion after verified evidence, reducing the risk of unfinished deployments and ensuring owner-controlled actions for sensitive tasks.

Workflow

• 1. Clone the agent-harness repository and run the installer. 2. Configure AGENTS.md and ~/.agents/skills to include the harness. 3. Define owner approvals in OWNER.md (e.g., money caps, account usage). 4. Issue a request to the agent. 5. The harness interprets the request and selects the appropriate agent role. 6. The agent performs the task and records each step in a ledger. 7. The done‑check.py hook verifies completion evidence. 8. If evidence is missing, the hook returns a NEEDS YOU message. 9. The agent continues or stops based on the verification result. 10. A scheduled cron or launchd job runs continue every few hours to keep the agent active. 11. Weekly reviews test harness changes against current tasks before adoption.

Evidence

The harness reads what you asked and picks how to act:

Source

Every row in a ledger is verified only with evidence: a test run, phone and desktop screenshots, or a URL check.

Source

NEEDS YOU (1 of 3): Stripe -> create the account -> unblocks checkout

Source