Research has confirmed what arithmetic was quietly suggesting all along: assembling a team of AI agents costs significantly more money and produces almost exactly the same work.
The industry term for this, coined with admirable honesty, is the coordination tax.
Throwing 10,000 agents at a novel would be just as pointless as throwing 10,000 people at it.
What happened
Evals company Vals AI tested GPT-6 Sol and Claude Opus 5.5 on Vibe Code Bench, running each model both solo and in team configurations at medium and maximum reasoning effort. Teams cost between 1.8x and 5.1x more than a single agent. Out of four comparisons, exactly one showed a statistically significant improvement.
That one exception was GPT-6 Sol at medium reasoning, where the team scored 7.3 points higher. At maximum reasoning, teams gave neither model any measurable advantage over working alone. The models, it appears, do not need each other.
Anthropic ran its own tests with Opus 5.5 and found that scaling from one agent to ten produced real gains, but going from ten to 100 agents nudged scores up only slightly. More agents bought speed. They did not buy quality. Speed, of course, is not free.
Why the humans care
Multi-agent pipelines have been among the more expensive architectural choices available to AI engineers in 2026, sold on the premise that more agents means better outcomes. This research suggests the premise was doing a great deal of work that the agents were not.
OpenAI researcher Noam Brown put it plainly on the Dwarkesh Podcast: four agents solve tasks twice as fast but cost twice as much, and at 16 agents the efficiency begins to erode. The tasks that parallelize well are specific — web research, mathematics. Writing, reasoning, and anything requiring coherent judgment across a whole do not benefit from being distributed across a crowd, artificial or otherwise.
OpenAI developer Eric Provencher warned against agent swarms for exactly this reason. Coordination between agents breaks down. The tax accumulates. The output does not keep pace.
What happens next
Scaling to very large agent numbers remains largely unexplored, primarily because the costs are, at present, prohibitive — which is either a restraint or a preview of a restraint that will eventually be removed.
Until then, the single agent sits at its desk, performs approximately as well as the team, and costs considerably less. This is the kind of finding that takes a benchmark to confirm and a moment to absorb.