The done rule: why an agent may not mark its own work complete

Four clauses. Verified end to end, as the real user, on the real devices, with the test data deleted. Each clause exists because of a specific failure.

1. VideoAI cannot mark work doneThe differentiator is not that an agent writes Apex. It is that the agent is not permitted to decide whether the work is finished.Preview8 min
2. VideoThe permission clause: as the real user, not as an admin7 min
3. VideoWhy an administrator’s successful test proves nothing about a standard userField-level security, sharing and record types all behave differently. A test run as an administrator is a test of the administrator’s access.7 min
4. VideoThe cleanup clause: create it, then delete all of it6 min
5. ReadingFour clauses, one rule, stated precisely5 min
6. QuizCheck: the rule, stated7 min

Field-level security, sharing and record types all behave differently by user. A test run as an administrator only tests the administrator’s access.

1. VideoField-level security, sharing and record types behave differently by user7 min
2. VideoMatching a permission set versus logging in as6 min
3. VideoOn the real devices, including the phone7 min
4. VideoCreate the test data, then delete all of it6 min
5. LabLab: run the same flow as an admin and as a standard user9 min

Sometimes there is no permission-scoped way to test a change. Saying so is not a failure of the work; calling it done anyway is.

1. VideoPending verification: the honest state7 min
2. VideoWhen a permission-scoped test is genuinely impossible6 min
3. ReadingSaying so, instead of calling it done5 min
4. LabLab: write a pending-verification note that says what is missing9 min
5. QuizCheck: as the real user8 min

Building the rule into a workflow so it cannot be skipped by a tired person at 6pm on a Friday.

1. VideoA workflow where the agent cannot set the done state itself7 min
2. VideoWho holds the done button, and why it is never the agent6 min
3. LabReviewing a completion claim, and finding the gap9 min
4. ReadingWhat “verified” has to include before a reviewer signs it5 min
5. QuizCheck: the workflow7 min

The best proof point in the portfolio: a fix, fully designed and quantified, that stayed unshipped until a human signed off on it.

1. VideoA fix, fully designed, deliberately not shipped“Not deployed. It changes the reported contract value on a customer-facing progress claim, so the controller should confirm first.”7 min
2. VideoWhy a progress-claim number needs the controller’s word, not the agent’s6 min
3. ReadingThe line between technically correct and safe to publish5 min
4. LabLab: write the confirmation request a controller can actually answer9 min
5. QuizCheck: the done rule9 min