AI Coding Agents Enter the Enterprise

The story of the past eighteen months in software is not that AI can write code — it could already do that — but that companies have started letting it write code unsupervised, inside their own repositories. Coding agents now open pull requests, run the test suite, read the failure, and try again. What began as an autocomplete feature in the editor has quietly become a colleague with commit access.

The adoption numbers are striking. Most large engineering organisations now report some form of agent usage, and a significant share of new code in those teams passes through a model before a human ever sees it. The figures vary wildly between surveys, and vendors have every reason to round up. What is not disputed is the direction: two years ago this was a pilot, and today it is a line item in the budget.

The practical change is in the shape of the work. Engineers spend less time typing and more time specifying, reviewing, and rejecting. Teams report that review queues have grown faster than anything else — when an agent can produce ten pull requests in an afternoon, the bottleneck moves to the humans who have to approve them. Several companies have responded by capping how many agent-authored changes can be open at once.

For startups, the economics are the headline. A four-person team can now ship what used to take twelve, which changes what investors expect before a seed round and what a founder can promise on a deadline. It also compresses the moat: if a competitor's feature took six weeks of engineering last year, assume it now takes two. Speed has stopped being a differentiator and has become the entry ticket.

The risks are the mundane kind, not the science-fiction kind. Agents introduce subtle bugs confidently, pull in dependencies nobody vetted, and produce code that passes the tests while missing the intent. Security teams worry less about a rogue agent than about sheer volume: more code, written faster, reviewed more thinly. The incidents reported so far look less like a robot uprising and more like an intern with too much access and no sleep.

Enterprises are converging on a familiar answer. Agents get scoped credentials, work on branches rather than main, and every change carries a label saying which model produced it. Regulated industries add a human sign-off before anything reaches production. None of this is novel — it is the same control set companies built for outsourced development, applied to a contractor that works at three in the morning and never asks a question.

So how do you talk about this in your own team without sounding like a brochure or a sceptic? Be specific about the task, not the technology. "We used an agent to migrate the test suite, and it took two days instead of two weeks" is a claim your manager can evaluate. "AI made us more productive" is not. If a number exists, use it; if it does not, say plainly what got easier and what got harder.

And be honest about your own role in the output. The engineers who come out of this well are not the ones who hide the tool, nor the ones who credit it for everything. They say what they delegated, what they checked, and what they caught. In a review, in a stand-up, or in a performance conversation, the sentence that still carries the most weight is the same one it always was: here is what I decided, and here is why.