Chapter C62 · Interactive Course

Check the work an agent returns

Diffs, matching a check to the risk, regression tests, failure output and shared blind spots.

What You’ll Learn in This Chapter

1. Review the actual diff

   Activity 1 of 7 · 5 min

   Inspect changed behavior in context.

   Inspect the actual patch directly, examining added lines, removals, and context. Verify how edits satisfy requirements without introducing side effects. Never accept an assistant summary without cross-checking the real git diff.

   Start activity →

2. Choose checks for the risk

   Activity 2 of 7 · 5 min

   Match a check to the behavior that changed.

   Choose checks based on what could fail. A calculation may need unit cases; a form may need interaction checks; a deployment setting needs environment-specific validation. More checks are not automatically better if they miss the changed behavior. Explain what each important check establishes and where its coverage ends.

   Start activity →

3. A regression test protects an earlier failure

   Activity 3 of 7 · 6 min

   Identify a test that fails before the fix and passes after.

   A regression test captures a behavior that failed so the same problem is detected if it returns. For a bug fix, it should meaningfully fail under the old faulty behavior and pass with the repair. Test the requirement rather than merely duplicating the implementation’s internal steps.

   Start activity →

4. Read failures before changing code

   Activity 4 of 7 · 7 min

   Use a failure report to select a hypothesis.

   Read the error output and relevant context before changing code in response to a failed check. The cause may be the patch, a missing dependency, a test assumption, or an unrelated existing failure. Reproduce and classify it, then make the smallest justified repair and rerun the affected check.

   Start activity →

5. Generated tests can share blind spots

   Activity 5 of 7 · 5 min

   Inspect test expectations independently of generated implementation.

   Generated implementation and generated tests can share the same mistaken assumption. A test passing proves only what its assertions actually check. Compare the cases with independent requirements, a known result, or the original reproducer. Include important boundaries the implementation might otherwise overlook.

   Start activity →

6. Human review closes the evidence gap

   Activity 6 of 7 · 4 min

   Identify what needs actual interaction or specialist judgment.

   Human review can assess requirement fit, usability, domain meaning, and risks that automated checks omit. Give the reviewer a concrete artifact and evidence they can inspect. Review is another source of judgment, not a guarantee; combine it with appropriate tests and runtime observations.

   Start activity →

7. Accept or return the patch

   Activity 7 of 7 · 6 min

   Review a supplied diff, test run and screenshot, then request one precise correction supported by the evidence.

   Start activity →