Twenty-three tool calls per task, forty-minute sessions, and a deny-list that ended up longer than our style guide. This is what changed when we stopped treating AI coding tools as autocomplete and started treating them as junior engineers with commit access.
On the first Monday of June we gave four coding agents write access to a staging branch, handed them a backlog of forty-one tickets nobody had prioritized in two quarters, and told the team to go get coffee. Not because we were confident. Because the backlog was the kind that only grows — the migrations nobody wants to own, the test coverage everyone agrees is a problem in the abstract, the small refactors that never beat a customer-facing feature in planning. If the agents made a mess, we’d lose a morning. If they didn’t, we’d learn something about how the next year of this team’s work was going to look.
Four weeks later the backlog was thirty-one tickets shorter, we had exactly zero production incidents traceable to an agent commit, and we had also written a deny-list longer than our actual engineering style guide. Here’s what that month taught us.
The shift nobody warned us about
The obvious prediction — agents write code faster — was true and also the least interesting thing that happened. What we didn’t expect was how much the shape of a sprint changed. A ticket used to move from “assigned” to “in review” in one jump. Now it moved from “assigned” through a loop: the agent picks up the ticket, works for anywhere from ten to forty minutes in a tool-call loop — reading the codebase, writing a failing test, fixing it, re-running the suite — and only then opens a draft PR with a written rationale. The engineer’s job stopped being “write the code” and became “read the rationale, check the diff against it, and decide whether the reasoning holds.” That’s a genuinely different skill, and not everyone was equally good at it on day one.
What actually shipped
1. The migration nobody wanted to own
We had a six-month-old ticket to move a legacy service off a deprecated internal auth library. It touched nineteen files and no single engineer wanted to spend a week on it for zero visible customer impact. An agent took four sessions across two days, and a senior engineer spent about ninety minutes total reviewing the diffs in chunks. It shipped with better test coverage than the code it replaced.
2. Test coverage that had been “next sprint” for a year
We pointed an agent at our three lowest-coverage modules with one instruction: don’t change behavior, just characterize it. It wrote 340 new test cases over the month. About 15% of them caught genuine, previously-unknown edge-case bugs — nothing severe, but real. The rest were exactly what we asked for: a safety net that didn’t exist before.
3. The PR nobody had time to write
Small ergonomic fixes — inconsistent error messages, a confusing config default, a missing loading state — are the first thing cut from every sprint because nothing breaks without them. They’re also, unsurprisingly, exactly the kind of self-contained, low-ambiguity work an agent handles well unsupervised. We shipped nineteen of these in the month. Individually trivial. Together, a visibly less annoying product.
The agent’s code was never the risk. The risk was an engineer approving a diff they hadn’t actually understood, because the diff looked plausible and the sprint was behind.
What we still don’t let it touch
None of these are technical limitations. The agents could do all of them. We decided they belong to a human, and a month in, we haven’t reversed a single one of these calls.
The part that actually worried people, and why it settled down
The engineers were not worried about being replaced. They were worried about being turned into full-time reviewers, which sounded like a demotion. What actually happened is that review got faster and more interesting — because the agent’s PR descriptions read like a colleague’s, not a diff dump — and the time that used to go into writing the boring 80% of tickets went into designing the next quarter’s architecture instead. Two engineers used the freed-up time to prototype a feature we’d wanted to build for a year and never had the bandwidth for. That’s not nothing.
We’re not running a fully autonomous sprint again next month — the deny-list stays, the human-approves-every-merge rule stays, and we’re not in a hurry to change either. But the backlog is thirty-one tickets shorter, and for the first time in a long while, nobody on the team is dreading the next planning meeting.