When More AI Agents Make Things Worse, and How to Avoid It

Every piece in this series so far has made the same basic case: independent AI perspectives catch things a single model misses. This one is the necessary check on that idea. A 2026 study out of the National University of Singapore and CUHK Shenzhen set out to test exactly when multi-agent AI actually delivers that promised diversity — and found a specific, common setup that produces the least diverse output of everything they tried.

The setup least likely to give you diverse ideas: assigning experts

Nuo Chen, Yicheng Tong, Yuzhe Yang, Yufei He, Xueyi Zhang, Qingyun Zou, Qian Wang, and Bingsheng He generated over 10,000 research proposals across 20 topics, comparing five ways of structuring an AI team: no defined roles at all (Naive), a senior expert directing junior agents (Leader-Led), a flat group of early-career personas with no hierarchy (Horizontal), a panel of distinct-domain experts (Interdisciplinary), and a deliberate mix of seniority levels (Vertical). They measured genuine semantic diversity with the Vendi Score, a metric that counts effective distinct idea-clusters rather than just counting outputs, validated against human judgment at 87% agreement.

"Junior-dominated horizontal collaboration achieves the highest diversity, and interdisciplinary expert teams the lowest, with only modest quality differences."

The actual numbers are the counterintuitive part:

• Horizontal (flat, no assigned expertise): Vendi Score 8.08 — the most diverse
• Vertical (mixed seniority): ~7.5
• Leader-Led (one authority figure): 5.6
• Interdisciplinary (distinct domain experts): 4.65 — the least diverse of all five

Giving agents named expertise — the exact move that sounds like it should widen the range of perspectives in the room — produced the narrowest one. Overall idea quality barely moved between conditions (7.88–8.50 on a 10-point scale), so this isn't a diversity-for-quality trade-off. The diversity was simply lost.

Why: agreement, not judgment, is the default

The paper's explanation is specific about the mechanism, and it isn't model incompetence:

"Although agents are prompted with distinct personas or roles to elicit diverse viewpoints, they remain grounded in shared inductive biases. […] interaction devolves into an Echo-Chamber Effect, where agents prioritize agreement over independent critique."

Crucially, they isolated why expert framing specifically hurts: it's not expertise on its own, it's expertise combined with implied authority. Under a genuinely flat topology, senior personas actually produced higher diversity than junior ones — ruling out seniority itself as the cause. It's the combination of distinct roles and a structure that lets one perspective act as the anchor everyone else measures themselves against. Junior agents in a Leader-Led setup didn't offer orthogonal critiques; they aligned their answers with the leader's.

The same collapse shows up with scale: adding more agents does increase raw diversity, but the amount each additional agent actually contributes — Vendi Score divided by agent count — falls from 1.03 at 3 agents to 0.47 at 7. New agents mostly restate territory the group has already covered rather than opening new ground. And stronger, more heavily aligned models made this worse, not better: the paper calls this the "Compute Efficiency Paradox" — better models produce higher-quality individual answers while converging on more similar content across agents.

What actually prevents it

The paper doesn't stop at diagnosis. It tested interventions, and two specific ones worked:

1. Independent generation before any discussion

A "blind-writing" phase — every agent producing its own answer before seeing anyone else's — started at the highest semantic diversity of any condition tested and held that advantage through the session. This technique has a name outside AI research too: Nominal Group Technique, a decades-old fix for exactly the human-brainstorming failure mode we covered in our piece on production blocking. The same structural fix that helps human groups turns out to help AI ones too, for a related but distinct reason — humans lose ideas waiting their turn; AI agents lose independence by reading each other's answers before forming their own.

2. Genuinely different models, not just different prompts

"Model heterogeneity rescues diversity in authority structures."

A mixed-model horizontal team outperformed every single-model baseline. The persona prompt ("you are an immunologist") is a much weaker source of genuine variation than the model itself — different labs, different training data, different alignment choices. Assigning five different job titles to five copies of the same model is a shallower kind of diversity than five actually different models answering the same question.

What this means in practice

This is the honest caveat to everything else in this series: putting multiple AI agents in a room does not automatically produce the independent thinking you're after. Two specific setups reliably kill it — a dense, shared-history discussion where everyone sees everyone else's answer before contributing their own, and any structure where one participant functions as the authority the others converge toward. Both are easy to build by accident.

The fix the paper itself points to lines up with how we'd suggest structuring a Room: broadcast a question to every model at once for the initial, divergent pass — genuinely independent, nobody anchored on anybody else's answer — and mix real providers rather than relying on one model playing multiple parts. Save the shared, cross-visible discussion for after independent answers already exist, and be deliberate about not letting any one participant function as the room's de facto authority during that divergent phase.

See our Best Practices guide for the concrete version of this — including why ModelSight's broadcast, with zero cross-visibility until you ask for a comparison, is structured the way it is.

Sources

• Chen, N., Tong, Y., Yang, Y., He, Y., Zhang, X., Zou, Q., Wang, Q., & He, B. (2026). Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling and Collective Failure in Open-Ended Idea Generation. arXiv:2604.18005.