An AI agent shipped a working tool while I slept. Here's what I did and didn't let it do

Last Tuesday night I gave a coding agent a job: build a command that runs long development loops against an approved spec, with a live dashboard so I can watch progress. Then I went to bed.

By morning it had produced 10 commits across 5 repositories, a working dashboard, 33 passing tests it had written itself, and a 12-page plain-English explainer of how the whole thing works.

That is the part that makes headlines. The part that matters is the fence I built around it.

What I let the agent do

Write code. Write its own tests. Run those tests and iterate on failures. Write documentation. Commit locally with its own name on every commit, so the history shows exactly who did what.

What I did not let it do

Push anything. Every commit stayed on my machine until I had reviewed the diffs the next morning. No exceptions, not even for documentation.

Merge anything. Merging is a human signature. It means “I stand behind this,” and an agent cannot stand behind anything.

Touch secrets. Every diff was scanned before review, and credentials live where the agent cannot reach them.

Declare success. An unused tool is not a result. The work only counted two days later, when the command ran a real task end to end: it shipped a data-pipeline change, 73 tests passing, merged after my review.

The moment that mattered

One moment from that real run tells you more than the overnight sprint does.

Mid-implementation, the agent found that my own spec contained a math error: I had estimated 87% data coverage where the true ceiling was 74%. It did not silently patch the spec to make its result look better. It stopped and asked.

That behaviour is worth more than raw speed, and it only happens if your setup rewards stopping.

The pattern, if you are trying this with your team

Let agents produce. Keep humans on the three verbs that carry accountability: push, merge, declare done. Log everything so the audit trail survives the enthusiasm. And measure the payoff on real work, not on the demo night.

One person plus agents is a real operating model in 2026. But the interesting design question was never “what can the agent do.”

It is “what do you refuse to delegate.”

Source post: Post | LinkedIn