AI spread the building around and left the fixing to a few.
At a company where I worked, AI became a craze overnight. They signed up for a pay-per-token account, gave access to the team's key technical people, and built up enormous expectations around it. There was no planning and no defined outcome. Nobody could say what to expect from it.
A month later, three large projects were solving the same problem, each driven by a different department, with people who didn't talk to one another. What was meant to solve an internal process problem turned into a fight between departments, with a lot of money spent, and landed on the CEO's desk. He put a puzzle together, combining a bit of what each group had built into a single system. Nobody was fully satisfied. I wasn't involved in the projects, and from day one I argued for planning, or at least for alignment and assigned responsibilities.
That case is a larger version of something I see on a smaller scale. AI spread the building around and left the fixing to a few. Lots of people started building at the same time, and the decision nobody made at the start ended up with the CEO.
Day to day, the path is shorter. Fixing what nobody identified falls to whoever knows the system end to end. At that company, the CEO put the puzzle together; on a team, it's the handful of people who know where each piece fits. It's the same phenomenon at two different levels.
Fixing always falls to the same people
I've been working with operations that traded planning for executing on ideas. The logic is seductive. If the agent produces in hours what used to take weeks, talking things through first seems wasteful: build it, show it, adjust it. Other things take planning's place, such as quality guardrails, very active communication about what's going on, and update meetings where people tell one another what they changed.
Even with these mechanisms, a lot of work ships with a problem nobody identified, and the team ends up fixing things it didn't even know were happening.
Imagine a team where someone asks the agent to simplify the date format in an internal API. The test passes, the guardrail doesn't complain, the change goes in, and it shows up in that week's update meeting as an improvement. At month-end close, a financial report that reads that field comes out with the day and month swapped. The person who notices is the only one on the team who knew the report depended on that field, and nobody ever decided this would be their job. They fix it at night, explain it in the channel, and go back to what they were doing. The fix never gets a task card, so it counts as an interruption. The next week it happens again, with something else, and again it's them, until their own project falls behind.
AI promised to spread the work around. It spread the building around, and the fixing ended up with a handful of people who know everything end to end, precisely the ones who were already overloaded before any agent showed up. They're the ones putting the puzzle together every day.
Mike Julian of Duckbill told Gergely Orosz that a team of five at one point had 60 open PRs, the equivalent of two days of code review alone. My bet is that queues like that tend to land in a few hands, and always the same ones.
The bottleneck is in the people too
In an engineering post, Anthropic, which supplies the agents, described going from around 2,500 merged changes a day in May 2026 to more than 30,000 in September. The bottleneck became the developer infrastructure built for human scale (source control, builds, and CI), and they built their own source control system. The authors say these ideas only make sense for enormous monorepos and that they don't expect many organizations to have this problem anytime soon. Sachin Malhotra, also at Anthropic, writes that the load on CI remains bursty because human engineers still drive and approve a large share of the PRs.
That's the scale of an AI lab, and on that point they're right. The mismatch between producing work and absorbing it also appears in ordinary operations, though, and in the ones I describe here it takes the form of overload on whoever does the fixing. When the tool couldn't handle the volume, Anthropic replaced the tool. In an ordinary operation, whoever can't handle the volume has a name.
Volume on its own is misleading. Orosz himself noted that Uber's AI power users generate 52% more PRs than developers who use AI less, and that the figure doesn't reveal the quality of those PRs.
Take a team where each person runs their own workstream with agents, and the update meeting becomes the only time everyone is in the same conversation. Each person says what they shipped. Almost nobody has time to think about how a colleague's release affects their own work. The person who knows everything end to end listens differently, connecting one comment to the next, and leaves the meeting with a list of things to check that nobody asked for. The update meeting worked. That person's queue grew too.
On a very senior team, maturity keeps much of this in check. On a less experienced team, I think the problem would be much bigger, because fewer people would recognize the damage before it got very far.
Even with maturity, managing has gotten more complicated. I'll admit I've had a harder and harder time keeping up with the operation. The rituals and controls have to change with every project, because the nature of what gets delivered changes, and the way the work is tracked has to be redesigned to match. Overseeing the work, motivating people, and keeping the team together have also become management challenges. Whoever does the fixing gets tired.
Speed is rented, judgment is in-house
The company pays for the speed agents provide, and that speed depends on commercial and technical conditions it doesn't control. What the company does own is the judgment that determines what gets delivered with that speed, and that judgment will be exercised somewhere.
That's why I argue for planning, or at least for agreeing on who does what: so more people exercise that judgment early, while sorting out the overlap still costs a meeting, sometimes a boring one, sometimes an inconclusive one. Without that conversation, the judgment happens anyway: later, at higher cost, and in the hands of a few. At the company in the story, in the hands of one. That holds even if the price of agents never goes up, because the cost of fixing late is paid in people, and nobody renegotiates that.
My bet is that for a reversible change, fixing pays off more than planning: a piece of on-screen text, a layout tweak, an internal script someone rewrites in the afternoon. And that the math flips when the change writes data, touches a contract with another system, or reaches the customer. At that point, fixing it means dealing with everything the error has already caused, and that date format turns into an investigation of which incorrect figures were sent out, and to whom.
There's a scenario that worries me more. If the cost of using agents goes up, commitments made at today's speed become a race against the clock. Deadlines set with cheap agents will have to be met with expensive agents or with people, and whoever has lost the habit of planning and coordinating will have to relearn it in a hurry, which is the worst way to learn anything. Whether it will happen, and when, I don't know. Teams that already coordinate before they start building will feel less of the impact.
Whoever skips the conversation at the start is choosing, without realizing it, who will put the puzzle together at the end.