Agent Practice Notes

What changes when the assistant can act

The speed is real, and so is the review it quietly skips.

• Commands run: installs, deletions, rewritten history.
• Twenty minutes of agent work outpaces a day of reading.
• The summary is written by the same thing that wrote the code.
• Nothing slows down enough for you to notice you stopped checking.

One task, one diff

The shape of a good session is the same whichever tool you use.

Small steps are not about the agent's ability to hold them. They are about yours to check the result.

• State one task with the finish line named.
• Ask for the plan before the edits.
• Let it work in a bounded step, not on an open mandate.
• Read the diff — every changed line, before anything else.
• Ask for proof: the output, not “should work now”.
• Commit small and often; commits are the checkpoints.

Where it helps and where it drifts

The characteristic failure is not nonsense. It is confident, well-formatted work that is wrong in one place.

The review step therefore cannot be delegated back to the thing being reviewed.

• Good at boilerplate, mechanical refactors and test scaffolding.
• Good at reading unfamiliar code and explaining it back.
• Weak on facts it cannot check in the repository.
• Weak on trade-offs nobody told it about.
• Weak at knowing when to stop and ask.