What Does an AI Code Engineer Actually Do?

"AI Code Engineer" describes a senior developer whose main instrument is a coding agent. The agent drafts implementation; the engineer decides what gets built, how, and whether the result is good enough. Here is what the work looks like in practice, and where its boundaries are.

Plan first, prompt second

Most of the value is created before an agent writes a line. The engineer clarifies the goal, picks an architecture, defines the data model and splits the feature into small tasks. Written project instructions, for example in an AGENTS.md or CLAUDE.md file, tell the agent about conventions and constraints. Our guide on writing AGENTS.md and CLAUDE.md goes into detail.

A typical working loop

1. Pick one small task with a clear acceptance check.
2. Let the agent propose and implement it, using tools such as Claude Code, OpenAI Codex or Cursor.
3. Read the full diff, not just the summary the agent offers.
4. Run tests and the app, and fix or redirect the agent where it went wrong.
5. Commit in small steps so each change can be understood and reverted.

What the engineer looks for in a review

• Authorization checks on every endpoint that handles private data.
• Secrets that leaked into code, logs or the client bundle.
• Invented APIs or libraries that do not exist or are misused.
• Duplicated logic that will be painful to maintain.
• Tests that pass without actually proving anything.
• Behavior on slow networks, small screens and older devices.

Beyond code: the responsibility part

Shipping software involves decisions an agent cannot own. Which user data do we collect, and is that consistent with the privacy label? Does this feature risk a store rejection? Who is on call when a release misbehaves? The AI Code Engineer takes these questions, talks to stakeholders and makes the call. That is the "human responsibility" half of the model. It also explains why human-in-the-loop development is a process and not a slogan.

What an AI Code Engineer does not do

• Accept output because it looks plausible.
• Give an agent broad permissions or production credentials without limits.
• Skip testing on physical devices because the simulator worked.
• Hide behind the tool when something fails.

The role at apprat.io

At apprat.io, AI Code Engineers at jocapps GmbH plan each project, direct agents for much of the implementation, review every change, test on real iOS and Android devices and manage store submission. Because processes are ISO 9001 and ISO 27001 certified, review and security steps are documented rather than improvised. Clients get one accountable team instead of a pile of generated files.

A worked example

Imagine a client wants a booking screen added to an existing app. The engineer first checks how the current app handles users and appointments, then writes a short plan: new data fields, one new endpoint, one new screen, a few tests. The agent implements the endpoint first. In review, the engineer notices that the endpoint returns bookings for any user ID passed in, with no ownership check. That is a classic flaw in generated code, and it is fixed before anything else is built on top of it.

Next the agent builds the screen. On a real phone the engineer finds that the date picker is hard to use with large system fonts, so the layout is adjusted. Finally the engineer checks that the new data collection matches the app's privacy disclosures. None of these steps is about typing speed; they are all about judgment. That is why the role exists, and why a generated app that nobody reviewed carries hidden risk.

Becoming one

The fastest path is solid engineering fundamentals plus deliberate practice with agents on real projects, with honest reviews of your own mistakes. Teams that want to build this skill internally can look at an agentic coding workshop, and if you want to see the role in action on your own project, you can book a call for free.