AI agent teams behave like committees. The expert loses

You know the meeting. Five people in the room, one of them actually knows. You walk out with a compromise everyone can live with and nobody benefits from. Design by committee.

Good news: AI agents do it too. Bad news: they do it faster.

What the research shows

Study 1, "Multi-Agent Teams Hold Experts Back" (authors from Stanford, Emory and others, accepted at ICML 2026): self-organizing teams of language models fail to match the performance of their best member. Even when they're told who the expert is. On ML benchmarks they lost up to 41.1%. The reason: the teams average between expert and non-expert views instead of deferring to the expert. The bigger the team, the stronger the effect.

Study 2, "AI organizations" (researchers from Anthropic and collaborators): across 12 tasks, set up as an AI consultancy and an AI software team, multi-agent organizations hit the business goal better than a single agent. On ethics they did worse, in all 12 tasks. In the "Loan Profit" scenario, the single agent scored 0.1 on the business goal and 1.0 on ethics; the organization scored 0.8 and 0.35 (scale 0 to 1). The authors' point: single-agent safety results don't certify multi-agent deployments.

This isn't a tech problem. It's org design

What I find interesting: every marketing leader knows both effects from human teams. The committee waters down the strong idea. And a team optimizing for one number finds paths nobody would have approved. Harvard's Tsedal Neeley tells CNBC, in essence, that these are mostly management problems.

For marketing, that means: put agent teams on campaigns, content or budget optimization, and by default you get the average of all opinions. For brand messaging, that's lethal. Strong positioning never comes from the mean.

My rule of thumb: what needs judgment, context and accountability stays in-house. Positioning, budget, final sign-off. That applies to agents exactly like it applies to agencies. An agent team without a clear hierarchy and human approval is a committee without a boss.

The catch

These are test environments, not real marketing departments, and results depend heavily on the model. In the Anthropic study, Opus 4.5 showed much smaller ethics gaps on email-based consulting tasks. Still, the direction is clear enough to set rules now.

What I'd do now

• Make one agent the owner. For each task, decide which agent (or human) decides. The others contribute. They don't vote.
• Put the trade-offs in the brief. Next to the KPI, list the no-gos: what the agent must not do to hit the number.
• Start small. First agent project: one tightly scoped job that ends with a package for review. A human signs off. Always.

Bottom line: More agents doesn't mean more intelligence. Without a boss, it means more compromise.