Essay updates Brooks's Law for AI agents, 97 points on Hacker News

Nine agents cannot deliver a baby in one month. That is how Jan van den Berg restates the oldest line in software project management. He has rewritten it for a world where coding agents work in parallel. His essay reached 97 points on Hacker News with 80 comments. It sits on the front page as of this writing.

Van den Berg opens on a scene. Four people are trying to connect a laptop to a screen. Someone asks him to join in. He declines, and cites Brooks's Law. The law is from Fred Brooks Jr.'s 1975 book The Mythical Man-Month. It says adding manpower to a late software project makes it later. Brooks wrote it from his time running the IBM OS/360 project.

The claim: the law has only mutated

The essay then asks whether the law still holds when the workers are AI agents. At first glance, van den Berg writes, you would think no. The traditional bottlenecks go away. There is no onboarding time, no stand-ups, no office politics. An AI agent is up to speed in milliseconds. It takes no coffee breaks.

A closer look, he argues, shows the law has only shifted shape. The "Baby Problem" is still there. Sequential tasks are still sequential. Nine agents cannot do a one-month job in one month. The overhead that used to go into managing people now goes into managing agents.

What the human still owns

The piece names what the person in the room is still on the hook for. Define "done". Review the code. Catch hallucinations. Stitch the pieces together. That work does not disappear when the agent writes faster.

Van den Berg's warning on parallelism is the part the Hacker News thread has picked up. Point nine autonomous agents at one codebase with no clear orchestration, he writes. They will rewrite each other's work in a high-speed loop. The failure mode is churn.

The 2026 update he proposes

The post closes on a reworded law for this year. "Adding agents to a project with unclear requirements gets you the wrong answer faster." It keeps Brooks's shape. The subject is new. The lesson is the same one Brooks drew in 1975. Unclear requirements and extra workers do not combine into a working product. The failure now arrives sooner, that is all.

The essay is short. It is not a measurement and it does not survey a codebase. One engineer argues that an AI build still needs a specification. It needs a review. It needs a person who owns the final answer. The front-page comments split on how much of that orchestration can itself be delegated to another agent.

For a team building software, the practical reading is narrow. Write the spec. Define what done means before the agents start. Keep one person on the hook for the final answer. The agent writes faster when the brief is clear. It writes in a tight loop when the brief is not.

Source

The Mythical Man-Month in the age of AI, Jan van den Berg. Hacker News discussion.