Brooks's Law for Agents: How to Scope Work So More Agents Help Instead of Hurt

A 1975 rule

In The Mythical Man-Month, Fred Brooks wrote one of the best-known lines in software: "Adding manpower to a late software project makes it later." New people need time to learn the system, some work can't be split, and every extra person adds conversations. Three people have three lines of communication. Eight people have twenty-eight.

For fifty years, that has been a reason to keep teams small and requirements clear. AI agents look like the way around it. They don't need onboarding, they don't get tired, and you can start five more in a minute.

Agents moved the cost

Agents did make one thing cheap: doing the work. Writing the code, drafting options and trying again cost very little now. What they didn't make cheap is deciding: knowing what to build, reviewing what came back and owning the result.

Every agent still reports to a person. Each new one adds output to check and another context switch for whoever checks it. If the task is clear, that is manageable. If it isn't, more agents produce more confident work in the wrong direction, and the reviewer becomes the bottleneck. As Jan van den Berg put it this week, adding agents to a project with unclear requirements gets you the wrong answer faster.

Four things we do before any agent touches a repo

• Map the real system. Before anything is generated, we read the code, data and workflow the agent will touch, inside the client's own environment. Most bad agent output starts with a wrong picture of the system.
• Write "done" as a check. Tests or acceptance criteria exist before the agent starts. If you can't say how you will know it worked, the task isn't ready.
• Draw the boundaries. What the agent can change, what is off-limits, and which actions need a person to approve them, enforced by permissions rather than only stated in the prompt.
• Size the work to the reviewer. We add agents only when tasks are truly independent and someone has the time to review each one. One well-scoped agent beats five that need untangling.

The scope card

We put all of this on one card per task, written before the first prompt: the task, what "done" means, what the agent can change, what is off-limits, what needs approval, who reviews it, and how many agents it needs. It takes minutes, and it is the cheapest way we know to stop agents from creating work instead of finishing it.

How we approach it at Zylen

Zylen is an AI optimization partner. Our forward-deployed engineers work inside the AI systems companies already run, and scoping comes first: measure what is there, then build. Our ten-day AI Optimization Audit maps where AI fits in your systems and what to fix first.

Scope first, then add agents: start the AI Optimization Audit →

Brooks quote from The Mythical Man-Month (1975).