Worker Agents, Specs, And Why AI Needs Boundaries

AI Agents Need Reviewable Boundaries

AI agents perform better when the task is bounded before work starts. Boundaries make the output reviewable: you can see whether the agent did the job or drifted into adjacent work.

Two Prompts, Two Risk Profiles

An AI agent can change many files quickly. That speed is useful when the task is clear and risky when it is not.

Compare two prompts:

Improve the plan page.

and:

Update the overdue bulk-reschedule menu so it supports tomorrow, next week, and custom date. Do not change unrelated task-list styling. Add tests for the new choices.

The second prompt creates reviewable work. The first creates a guessing game.

This is standard engineering discipline applied to AI: define the issue, constrain the scope, inspect the diff, and run the checks.

What To Define Before Delegating

Before delegating to AI, define:

• goal
• files or area of ownership
• what is out of scope
• acceptance criteria
• verification command

This is not bureaucracy. It is how you make fast work safe.

Better AI Work Prompts

Weak:

Make the app better.

Better:

Reduce duplicate network loads in this provider by using the shared dedupe helper and add a StrictMode test.

Practical Takeaway

The best AI task is small enough that a human can review it carefully. If the desired change touches multiple product flows, split it into specs before delegating.

Fast generation does not remove the need for ownership.

Common Mistakes

Do not confuse output volume with progress. A small diff that solves the right problem is better than a large diff that changes the product direction.

Use FocusToday

In FocusToday, hold AI work as normal tasks: one bounded goal, one active task, and a clear review step before calling the work done.

Sources And Influences

Influenced by software engineering review practice, Getting Things Done, and project management principles.