Nature Study Finds AI Agent Teams Can Cut Sequential Planning Performance by 70%

Adding AI agents did not reliably improve results in a controlled study across six benchmarks. On sequential planning, multi-agent systems reduced performance by as much as 70%, underscoring that task structure and coordination design shape the outcome.

The peer-reviewed study, Capable language models can outgrow the benefits of collaboration, compared a single-agent system with four multi-agent designs. It held prompts, tools and total compute budgets constant across 260 configurations, six agent benchmarks and models from three major language-model families.

The four team designs used centralized, decentralized, independent and hybrid coordination. That setup was intended to isolate how systems were organized, rather than give one design more resources than another.

Parallel work and step-by-step work diverged

The sharpest contrast came from task design. On a financial-analysis benchmark, the strongest multi-agent system improved performance by about 81%. The researchers linked that gain to agents investigating relatively independent information in parallel; sequential planning imposed dependencies for which collaboration became costly.

Two benchmarks, opposite results

01+81%

Financial analysis

The best multi-agent system improved performance by about 81% on the financial-analysis benchmark.

02−70%

Sequential planning

Multi-agent systems reduced performance by as much as 70% on the sequential-planning benchmark.

Coordination is a cost, not a free upgrade

The researchers identify communication overhead, repeated exchanges, synchronization problems and error propagation as costs that can erase a team’s advantage. Their trace analysis found that centralized-verification designs contained error propagation better than systems where agents worked independently.

Matched compute budgets did not mean the systems took the same number of interaction steps. Hybrid systems used 6.2 times as many turns as a single-agent system; centralized and decentralized systems used 3.8 and 3.6 times as many turns, respectively.

A selection rule with clear limits

Single-agent baseline performance was the most reliable indicator of whether coordination would help. Once that baseline exceeded roughly 45%, additional agents were generally less likely to improve results. The rule predicted the direction of the multi-agent effect in 94% of validation configurations, but the authors call it a practical rule for these settings, not a universal scaling law.

A predictive model selected the best architecture in 87% of held-out configurations within the studied domains. Whether that approach transfers to other domains or teams combining models with genuinely different capabilities remains a question for further study.