Why I Built Inkspect: How Do I Know the AI Is Really Done?

Description and use

The author built Inkspect to verify that an AI agent truly completes a task by reviewing findings, evidence, and fixes before moving on.

Problem

When a coding agent claims completion, the output may still contain hidden issues that surface later, causing regressions and wasted effort.

Outcome

After integrating Inkspect, the team could detect that interrupted runs lost model IDs, add them to checkpoints, and confirm the fix with passing tests for failure, cancellation, and recovery scenarios.

Workflow

• 1. Assign a task to the coding agent.
• 2. Agent completes the task and generates a report.
• 3. Inkspect reviews the report, identifies findings, and records evidence.
• 4. If needed, Inkspect applies fixes and re‑checks the code.
• 5. Once all findings are resolved, the agent receives a handoff to start the next task.

Evidence

An agent would finish a task, but more problems would appear when I checked the result.

Source

The engine recorded model IDs returned by the AI service. These labels help explain which models were involved in a review.

Source

The fix was to include the session’s recorded model IDs in each stage checkpoint.

Source

The test checked three cases. The saved test results show the difference: ... After the fix, all cases passed.

Source