Engineering with AI Without Losing Control

The interesting question about coding agents is no longer whether they can produce code. It is whether a team can turn that capability into reliable delivery.

Give the agent a bounded job#

“Improve the service” is not a workable instruction. A strong task identifies the behavior to change, the repository and branch, the constraints that must remain true, and the evidence required for completion.

This is familiar engineering discipline. AI simply makes vague requirements fail faster and at a larger scale.

Separate discovery from mutation#

For unfamiliar systems, the first pass should be read-only: trace the active code path, configuration, deployment state, and data ownership. Only then should the workflow propose changes.

That separation is particularly important around production systems. A plausible guess is not evidence, and an automated write is still a production write.

Treat verification as a deliverable#

The agent should return more than a diff. Useful evidence includes:

• focused tests for the behavior that changed;
• a full build or type check when appropriate;
• a visual check for user-facing work;
• deployment health and a production read-back;
• explicit notes about anything not verified.

The reviewer can then evaluate a compact chain of evidence instead of reconstructing the entire session.

Keep human gates where consequences change#

Some boundaries deserve confirmation: writing production data, publishing externally, granting new access, or changing infrastructure shared by other services. The goal is not to insert approval into every command. It is to make the moment of increased consequence obvious.

The best AI workflow is not autonomous everywhere. It is predictable everywhere.