Brooks’s Law for AI Agents Still Has a Coordination Cost

Brooks’s Law for AI agents is a useful warning: adding agents can speed up some work, but it does not make coordination overhead disappear. The source article argues that the bottleneck has shifted from coordinating people to orchestrating agents—not that the underlying constraints have gone away.

Brooks’s Law for AI agents still applies to sequential work

AI agents can reduce onboarding delays, but they cannot make a sequential task parallel. The article’s shorthand is that nine agents still can’t deliver a baby in one month. More workers do not shorten work that depends on an earlier step being finished first.

That matters when planning an agent workflow. Splitting a task into more assignments only helps when those pieces can genuinely proceed independently. If each agent needs the same decision, artifact, or result from another, the added capacity may just create more handoffs to manage.

Code review and integration remain on the critical path

People still have to define what completion means, review code, catch hallucinations, and integrate agent contributions. Those responsibilities are not incidental cleanup; they are part of getting from generated output to a working result. Agents can produce contributions faster than a team can safely assess and combine them.

The integration risk gets concrete when agents work on the same codebase without clear orchestration: they may overwrite each other’s work. That is not a reason to avoid multi-agent setups. It is a reason to treat ownership and integration as design problems rather than assume parallel output will assemble itself.

My read is that review capacity is the limit worth measuring, not the number of agents you can start. A workflow that produces more code than anyone can check has increased throughput only on the output side. The useful question is whether the work arrives in pieces that can be reviewed and integrated without creating a new queue.

Add agents where the work can absorb them

The source article’s warning is direct: adding agents to a project with unclear requirements gets you the wrong answer faster. In practice, that means agent count is a poor proxy for progress. Clear completion criteria help people judge whether a contribution is usable; oversight is still needed to catch errors and make the pieces fit.

I’d add agents to work that can absorb parallel contributions, while keeping the human review and integration path visible. For a task with dependencies, unclear acceptance criteria, or shared files, adding more agents may move the bottleneck rather than remove it. The decision is not “agents or no agents”; it is whether the workflow has enough independent work to justify the extra coordination overhead.

A practical test: compare the time saved on parallel work with the time spent reviewing and integrating its output; if the latter grows faster, stop adding agents.