Speed changes the bottleneck

An agent can read a repository and produce a plausible patch quickly. Plausible is useful, but it is not yet correct. The hard part becomes deciding what should change and noticing when the result only looks finished.

That makes a precise brief worth more. A good task describes the behaviour you should be able to observe, the boundaries, and what must stay untouched.

Ask for evidence

I want a test to fail for the missing behaviour before the implementation exists, then pass for the right reason. For interface work, I also want the rendered page measured and inspected at the sizes people will actually use.

Diffs, tests, browser state and build output give a review something to check. A confident summary without them is just a claim.

Keep responsibility human

The agent can do more of the typing, but it cannot own the consequences. Checking the assumptions, the failure modes and the end result is still the engineering work.