When a delivery date slips, the offer on the table now is to start ten more agents tonight. Tamas Mezei, writing on brokenrobot.xyz on 14 October, says that offer fails for the reason adding people always did. His essay borrows its title from Fred Brooks’s 1975 book The Mythical Man-Month, and he quotes Brooks’s law as Brooks stated it: “Adding manpower to a late software project makes it later.”
Mezei’s thesis is that cheap code from AI shifts the real constraint onto deciding, verifying and coordinating, work that was always mostly organisational. AI then amplifies whatever coordination a company already has. This is reasoning, not a new measurement. Mezei reports no data from his own team or employer. The case rests on other people’s research and experiments. He concedes that the wider AI studies he cites tell us about AI use broadly, and say nothing direct about fleets of agents.
He is careful with Brooks. He notes that the law concerns late projects, whereas agents raise a question of throughput, and that Brooks later called it a rule of thumb. What Mezei carries over is the mechanism, in three parts. Work with sequential steps does not shorten when more hands arrive. The Google Research study he cites used small teams, three or four agents each, and was published in January 2026. Sequential planning got 39 to 70 percent worse under every multi-agent configuration tried, while a task that could be divided cleanly gained about 80 percent.
The second part is communication. Cursor began with agents coordinating as equals, and its engineer Wilson Lin reported that a group of twenty ended up delivering what two or three would. The third part is training: an agent takes in a codebase much more quickly than a new hire does, but someone still has to write down the goals and plan and keep them current.
Mezei raises the obvious rebuttal himself. Cursor fixed its problem with planner and worker agents, and tools for merge queues and review agents keep arriving. His answer is that those fixes are Brooks’s own ideas: the planner is the surgical team, with one chief programmer making the key decisions, and the single design authority is conceptual integrity. They settle coordination among the agents. They stop short of the people deciding what the agents are pointed at: the definition of done, the goal and the trade-offs.
The concrete symptom, in his telling, is a queue. Weak coordination used to feel like a slow team. Now it shows up as pull requests awaiting review, features nobody asked for, and several agents solving one problem in different ways. Review loops were tuned to human speed, he argues, and nothing adjusts them automatically.
His organisation-level evidence is mixed, and he says so. Faros AI drew on data covering over 10,000 developers and found engineers finishing twice as many code changes while company metrics stayed flat. In its 2026 dataset, the median pull request spent 441 percent longer in review. The DORA research programme found a positive link between AI adoption and throughput but a negative one with delivery stability. Both are correlational, and they disagree.
Unlike many diagnoses, this one comes with a prescription. Mezei proposes judging AI work by what customers do rather than by lines or pull requests. He cites Microsoft’s finding that only roughly one in three of the ideas it tested moved the metric they targeted. He wants shared, written intent with named owners before the number of agents grows, and review capacity that keeps pace with how fast output now arrives.
Our own earlier reporting points the same way, though it is separate from his evidence. A Vals AI study found teams of AI coders cost up to 5.1 times more and rarely scored higher, and Anthropic’s data showed quality plateauing beyond ten agents while token costs kept climbing. Taken together, these suggest a rule of thumb the essay implies but does not state: an organisation that could not coordinate ten engineers is unlikely to coordinate a hundred agents.
Before approving a bigger fleet, then, ask whether the team can write one page saying what the agents should build, for whom, and who decides when it is done. If nobody can, the first thing the extra agents will produce is a longer review queue.
Based on Tamas Mezei’s essay “More Agents, Same Mythical Month,” published on brokenrobot.xyz on 14 October 2026.
