Shipping a Field App with AI Coding Agents

Most of the code in our field operations app was written by AI coding agents. None of it shipped because an agent said it was done. This note is about the part in between: how we set up the work so agents can move fast without the app getting worse.

Why agents, and why now

A field app has a long tail of small, exact behaviour. A visit that repeats every second Tuesday has to land on the right day across a daylight-saving change. A crew member removed from a team has to stop receiving that team's calls at once, not at the next sign-in. Each rule is simple. There are hundreds of them, and they interact.

That is the kind of work current coding agents do well when the target is precise, and badly when it is vague. So most of our engineering effort goes into making the target precise.

Specs before code

Every piece of work starts as a short written spec: what the screen or service does, what it must never do, and the cases we already know are hard. Agents work from the spec, not from a chat. When an agent finds a gap, the spec changes first and the code second, so the next agent doesn't rediscover the same question.

The suite keeps everyone honest

The app has an automated suite of more than 1,600 tests, and it runs on every change. Agents write tests alongside features, but the tests that matter most are the ones written from the spec before the feature exists: they describe the behaviour we want, and the agent's work isn't finished until they pass.

When a bug reaches a phone, the fix starts with a failing test that reproduces it. That test stays forever. Over time the suite becomes a record of every mistake the app is no longer allowed to make.

Reviews by agents, decisions by engineers

Large changes get a second pass from separate review agents, each looking for one kind of problem: data that could be lost, a permission that could leak, a phone that could ring for the wrong person. Their findings go to an engineer, who decides what is real. Agents are good at finding candidates; people are still better at judging which ones matter.

What we've learned

Agents make the cost of writing code small, which moves the cost to specifying and checking it. The teams that get the most from AI coding are the ones that invest in that second half: clear specs, a suite that runs on everything, and reviews aimed at the failures that hurt users most.