A faster specialist can still spend the afternoon waiting for someone else’s answer. The same problem belongs in an assessment of AI agents.

In an October 29 DeepMind Institute essay, Alex Imas, Andrew Koh and Shane Legg separate two capabilities: doing a task and coordinating the people or agents who do the work. Their economic framework argues that improving the first can raise productivity while leaving firms broadly recognizable. Improving coordination past a critical threshold could change how large firms can become and how markets organize.

The mechanism is division of labor. A team can combine complementary strengths and move work between specialists. Copying one model adds capacity, but copies may share the same weaknesses. A strong individual score therefore leaves open whether a group can divide an unfamiliar problem, exchange useful context and catch mistakes before they spread.

These are conditional economic scenarios, not evidence that an agent organization has crossed such a threshold. The essay itself says today’s AI systems coordinate poorly in practice. The Institute also identifies its pieces as authors’ ideas and research, rather than Google’s official position. A forecast about concentration or wages should retain those qualifications.

For an enterprise, the useful next step is a workflow comparison. Give a single agent and a small agent team the same bounded job, tools and acceptance criteria. Keep a record of duplicated effort, lost context, correction work, elapsed time and the result a person actually accepts. Then increase the team size and see whether useful output grows faster than the coordination burden.

That comparison can expose a very ordinary bottleneck: several capable agents waiting, repeating or repairing one another’s work. Before buying a larger team, find out whether it can hand over a job cleanly.