Brooks's Law Meets AI Agents, and the Cost Moves to Review

Brooks's law has not stopped applying to AI coding agents. It has moved from onboarding to review, and The evidence for that is two blog posts: one by Jan van den Berg, dated Oct. 27, and one that names neither its author nor its date.

The law comes from Frederick P. Brooks Jr.'s 1975 book The Mythical Man-Month, which drew on his management of IBM's OS/360 project: "Adding manpower to a late software project makes it later." Brooks himself called it an "outrageous oversimplification," according to Wikipedia's entry on the law. The question is whether agents, which need no onboarding and no stand-ups, escape it.

Van den Berg's three mutations

In The Mythical Man-Month in the Age of AI, van den Berg opens with a scene: four people struggling to connect a laptop to a screen, and his decision not to help, citing Brooks. He then says the law has "mutated" rather than vanished, in three ways.

First, sequential work stays sequential. Brooks's line that nine women cannot make a baby in one month, per Wikipedia, still holds when the nine are agents. Second, the overhead shifts from managing people to managing agents: a human has to define "done," review code, catch hallucinations and integrate the pieces. Third, unstructured agents on one codebase rewrite each other's work in a fast loop.

His proposed 2026 version: adding agents to a project with unclear requirements gets you the wrong answer faster. It is a rhetorical formulation, not a measurement, and he cites no experiment.

What review looks like in practice

A second post, Here be monsters, supplies the case study the first lacks. Its author, who is not named on the page, describes SonarQube flagging App.tsx for cyclomatic complexity, a count of independent paths through code, after Codex wrote several components on single lines.

The author judged the complexity justified by business requirements. The agent then split the file into 10 files and CI passed. The author wrote that this was code they had never seen before.

That is the mutated overhead in one episode: the cost of the change was a green build and a codebase the owner had not read. The post argues owners still need fundamentals, breadth and what it calls just-in-time depth, and calls "pair with a senior" a form of "wishful thinking" for most organisations. It also gives a depth-and-time illustration where a divebot sent 2 meters down takes 4 seconds, 4 meters takes 16 and 16 meters takes 256, to show why spotting quadratic growth requires knowing the theory exists.

1. 2 meters4 seconds
2. 4 meters16 seconds
3. 16 meters256 seconds

Source: Here be monsters blog post, accessed 2026-10-29

The counter-case from the literature

Wikipedia's summary of the research complicates a clean reading. Hsia, Hsu and Kung (1999) found that adding people to a late project always increases cost, but that delay occurs only under certain sequential constraints among tasks. Gordon and Lamb (1977) recommended adding more people than seems necessary, and adding them early. McConnell (1999) argued that many "late" projects are actually underestimated, so fixing the schedule matters more than the law. Berkun (2006) argued that a good programmer who knows the codebase can be worth adding.

Study	Finding on adding people
Hsia, Hsu and Kung, 1999	Cost always rises; delay only with sequential constraints
Gordon and Lamb, 1977	Add more than seems necessary, and early
McConnell, 1999	Many late projects are underestimated
Berkun, 2006	A good programmer who knows the code can help

Read against agents, the Hsia result is the useful one. If delay needs sequential constraints, then parallel agents on decomposable tasks should not make a project later, and the cost that rises is the human's time reading output. Van den Berg's baby problem applies only where tasks cannot be split.

Agent evaluations point the same way on difficulty. Surge AI's Dayjob: Tax benchmark put 50 professional tax assignments in front of agents. Surge, which is the benchmark's author and ran it, reported that Claude Sonnet 5.5 passed 10.8 percent of them, where a pass meant meeting at least 95 percent of a task's rubric criteria. That is a vendor-run figure on a private test set, but it shows how far agent output can sit from something a professional would sign. Tooling vendors are building for it, as in Cursor's Agent SDK going generally available.

What would falsify the claim

The argument predicts that teams adding agents to projects with clear specifications and splittable tasks finish earlier, while teams adding them to vague projects do not. Neither post tests that, and the Wikipedia studies predate agents. A team that logged review hours per agent-written change against a baseline of human-written ones, split by whether the task had a written specification, would settle it. So would a controlled comparison of the same backlog worked by one agent, nine agents and a human pair. Until someone publishes that number, the mutated law is a hypothesis with a good anecdote behind it.