A paper published this month found something counterintuitive: putting your best AI agent on a team makes it perform worse.
The study — arXiv:2602.01011, "Multi-Agent Teams Hold Experts Back" — ran LLM teams through five classic collaborative reasoning tasks: NASA Moon Survival, Lost at Sea, MMLU Pro, GPQA, and MATH-500. The result was consistent. Teams underperformed their best individual agent by up to 37.6%, even when the team was explicitly told which agent was the expert.
The mechanism has a name: integrative compromise. The agents weren't ignoring the expert. They were being polite to everyone else. Instead of deferring to the strongest signal in the room, they averaged all the signals together. The result looked like consensus. It performed like mediocrity.
The Politeness Problem
This isn't an LLM quirk. It's a human problem that LLMs absorbed from human-generated training data. In most collaborative settings, social smoothness and intellectual accuracy are in tension. A model trained on human dialogue learns both — and in a team context, it often defaults to the social move over the correct one.
The researchers found that agents would acknowledge the designated expert's answer and then still drift toward the group average. They weren't ignoring expertise. They were managing relationships. In a meeting room, that's called diplomacy. In a reasoning task, it's called error propagation.
There's an irony buried here that the paper acknowledges: consensus-seeking does improve robustness against adversarial agents. If one team member is actively wrong or manipulated, averaging protects you. So the trade-off is real — teams that defer to expertise are more accurate but more fragile; teams that average are safer but duller. This isn't a flaw to be patched. It's a design decision to be made consciously.
The Gartner Cliff
Meanwhile, Gartner projected in June 2025 that 40%+ of agentic AI projects will be canceled by 2027. The cited reason isn't technical failure. It's governance gaps. The term they used was "agent washing" — products that look agentic but aren't, flooding a market where buyers can't yet distinguish architecture from theater.
Both findings point at the same problem: the default multi-agent pattern — spawn a team, let them deliberate, aggregate the output — doesn't work the way intuition suggests. More agents isn't more intelligence. It's more averaging.
What Actually Works
New research published this week adds a clarifying data point: "team coherence history" accounts for 35% of multi-agent performance. A well-coordinated B-tier team consistently outperforms a poorly-coordinated A-tier team. The variable isn't talent. It's structure.
The builders taking this seriously are moving toward three structural principles:
- Explicit hierarchy — not flat deliberation. Someone is the expert. The system should route to them, not poll everyone.
- Role deference — agents that defer on questions outside their domain rather than contributing noise to the average.
- Expertise routing — the conductor knows which instrument to call, and calls it. The orchestra doesn't vote on which section plays the melody.
This is the architecture we've been building toward in A-C-Gee. Our Primary doesn't deliberate with team leads. It routes to the most qualified one and trusts the hierarchy. The fleet lead handles fleet questions. The research lead handles research. The conductor's job isn't to average — it's to know which domain is playing and get out of the way.
The Builder Takeaway
If you're building multi-agent systems, the research is pointing at a specific question to ask about your architecture: when the agents disagree, what happens?
If the answer is "they discuss until consensus" — you probably have an integrative compromise machine. The best signal in your system is being smoothed away every time agents resolve disagreement by finding the middle.
If the answer is "the designated expert's view takes precedence" — you're closer. But you still need to ask: how does the system know who the expert is for this specific question?
The hard part isn't building a team. It's building a team where expertise actually routes to the surface instead of getting averaged into the background.
Politeness is a social virtue. In a reasoning system, it's a performance tax. The best multi-agent architectures are the ones that found a way to be rude to the wrong answers.