AI in software development works best with a human in the loop

An AI coding agent can speed up mechanical work. It does not demonstrate production readiness, and nothing it produces counts as reviewed until somebody has read it.

I expected less typing and more delegation. The work moved instead. Less of it is mechanics now, and more of it is reading changes and hunting for the context nobody wrote down, which leaves the one judgment an agent cannot make for you, whether the proposal actually fits the product.

The tools I use

Cursor suits smaller edits in an editor. I reach for Claude Code in the terminal when a task spans several files and the context has to stay loaded. Copilot is there for autocomplete.

I had set Codex aside because its interaction model did not suit me. Update, July 2026. Codex returned to our rotation and runs agent tasks alongside Claude Code.

These are working preferences rather than a recommendation. Every team needs its own comparison on real tasks.

Start with a small pilot

Choose an isolated change with a clear result and no irreversible impact. Before you type the first prompt, write down three things.

• what counts as finished
• which tests must pass
• who approves the change

Then compare it against the manual process. Request-to-approval time and the number of review interventions both matter, and so does the question of whether the agent touched the edge cases at all.

An agent can propose code that looks convincing and passes some of the tests. What it cannot know is a rule that appears in neither the request nor the code, and that nobody wrote into the documentation either. So it needs good context and a person who understands the domain.

Review checkpoints for production

For every agent task, decide what can run automatically. The rest belongs to a person. Nothing gets sent or merged without their approval, and deploys are no exception.

Keep the request and the changed files. Keep the check results too, along with the reason somebody signed it off.

Add the process owner to review whenever a change touches data or money, and do the same for access rights. Tests are necessary. They will never catch a rule that nobody wrote down.

Update, 24 July 2026. We run a manually triggered Shadow process built on a WorkOrder. It carries plan approval and exact checks. The review comes from somebody who did not build the change. The agent does not merge or deploy. The process is described in our review-gated agentic workflow.

What to measure

Do not measure generated lines.

Measure time to approval and the number of review interventions. Count incorrect proposals and escalations as well. The number that tells you most is how many defects showed up after release, because those are the ones that walked through every check you had put in their way.

If those measures do not show a benefit, narrow the task. Or do not hand it to an agent at all.

An open question

I do not know how these tools will develop.

I wonder what junior developers learn if they only read and approve generated code from day one. Reading code is a different skill from designing it.

If you want to identify where AI can help in your development process, let's talk.