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

An AI coding agent can speed up mechanical work. Without human review, it does not demonstrate production readiness.

I expected less typing and more delegation. In practice the work moved. I spend less time on mechanics and more on reading changes, finding missing context, and deciding whether a proposal fits the product.

The tools I use

Cursor suits smaller edits in an editor. I use Claude Code for larger terminal tasks when several files and longer context matter. I use Copilot mainly 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, not 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 starting, write down what must be complete, which tests must pass, and who approves the change. Compare request-to-approval time, review findings, and edge-case coverage with the manual process.

An agent can propose code that looks convincing and passes some tests. It does not know a rule absent from the request, code, or documentation. That is why it needs good context and a person who understands the domain.

Review checkpoints for production

For every agent task, define what can run automatically and what a person must approve before sending, merging, or deploying. Keep the request, changed files, check results, and reason for approval.

For changes affecting data, money, or access rights, add the process owner to review. Tests are necessary, but they do not catch a rule that was never written down.

Update, 24 July 2026. We use a manually triggered Shadow process with a WorkOrder, plan approval, exact checks, and independent review. 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, review interventions, incorrect proposals, escalations, and defects after release. If those measures do not show a benefit, narrow the task or do not delegate it to an agent.

An open question

I do not know how these tools will develop. The question remains what junior developers learn if they only read and approve other people's code from the start. Reading code is a different skill from designing it from scratch.

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