Execute, Amend, or Queue
Explore how to manage AI agent work within defined boundaries by using Scope Contracts that specify outcomes, expected files, exclusions, and evidence. Learn to decide when to execute, amend, or queue discovered tasks to prevent scope drift and verify completion. This lesson helps you maintain task focus and reliability across sessions by controlling what agents can change and ensuring each change is validated.
We'll cover the following...
- Why did reasonable changes become scope drift?
- Isn’t fixing nearby code more efficient?
- What gives the task a real boundary?
- How small should one active outcome be?
- How do you define the expected surface?
- What belongs in the exclusions?
- How should the harness handle discoveries?
- What proves the work is finished?
- Can you stop drift while the session is still running?
- What can you add to your harness?
- How do you test the contract?
- What’s next?
Wrenfold’s environment passed the readiness gate we built earlier. The next ticket was CLIN-511: cancel an appointment and book its replacement without losing the original slot when the replacement fails.
The agent opened src/routes/appointments.ts, noticed inconsistent cancellation names, and started cleaning them up. Each fix exposed the next one.
Two hours later, eleven files had changed. The cancel-and-rebook path still released the original appointment before it knew whether the replacement was valid.
Every step answered something the agent had just read. Spend the same session yourself below. You get eight steps, and each one buys either the work the ticket funded or the thing the agent just noticed.
Spend the steps on what the agent noticed and the arithmetic comes out the same way every time. More work, less of the requested outcome.
Why did reasonable changes become scope drift?
Each change made sense in isolation. ...