2026-09-05
Across 260 matched-compute setups, extra agents rarely help past a ~45% single-agent baseline; finance can gain 81%, sequential planning can drop 70%.
Slogans like "more agents is all you need" treat multi-agent coordination as a free upgrade. In practice, prompts, tools, and compute usually move together, so a gain can be architecture or just a bigger bill. Google Research, MIT, and Google DeepMind pin those three constants and vary only topology and model tier. The question is blunt: when does an extra agent help, and when does it get in the way.
They restrict the claim to agentic work: multi-step interaction with an environment, partial observability, and strategy updates from feedback. One-shot coding quizzes do not count. Five-agent teams that hit 89% on HumanEval can look very different once the job is iterative debugging and repo navigation, where coordination tax and error cascades take over.
The template is strict. Same task prompts, same tools, same per-system reasoning-token ceiling; only coordination structure and model capability change. Five topologies form a structural ablation: a single agent (SAS); independent parallel workers plus a synthesizer; decentralized peer debate and voting; a centralized orchestrator; and a hybrid of orchestrator plus peer chat. Models span OpenAI (GPT-5 nano/mini/full), Google (Gemini 2.0 Flash, 2.5 Flash/Pro), and Anthropic (Claude Sonnet 3.7/4/4.5), with intelligence-index values from about 42 to 71. Six benchmarks cover web browsing (BrowseComp-Plus), financial analysis, game planning (PlanCraft), workplace tasks (WorkBench), software engineering (SWE-bench Verified), and CLI work (Terminal-Bench). N = 260 configurations. Mean reasoning budget is about 4,800 tokens.
They also score process, not only final accuracy: coordination efficiency, overhead, error amplification, message density, redundancy, plus the single-agent baseline, tool count, and team size. A linear model with prespecified interactions is fit on those features. Capability is either an external intelligence index or a task-grounded Agentic Capability Index (ACI): the model's mean SAS score across the six benchmarks.
Averaged over all six benchmarks, MAS improvement versus SAS is 0.0% (95% CI −58.7% to 77.2%, σ = 37.5%). Sign and size follow decomposability.
Finance Agent is the friendly case. SAS mean 0.349; centralized 0.631 (+80.8%); decentralized +74.5%; hybrid +73.1%. A single agent searches news, then filings, and never fans out; three specialists cover regulation, SEC documents, and operations, and the orchestrator merges. PlanCraft is the hostile case. Crafting a diorite wall is three steps for SAS. MAS invents fake subtasks (look up the recipe, check inventory, then craft) and burns the budget on messages. Independent drops −70.0% (0.170 vs 0.568). WorkBench tops out at +5.6%. On BrowseComp-Plus, decentralized is +9.2% and independent is −35%. On SWE-bench Verified, SAS already sits at 0.488; every MAS variant slips, from hybrid −1.3% to independent −12.8%. Terminal-Bench SAS is 0.312; independent +6.0%, centralized −20.0%, because two tools do not justify a heavy orchestrator.
The predictive model reaches cross-validated R² = 0.373 with the intelligence index and 0.413 with ACI. Within domain it picks the best architecture on 87% of held-out configs (20% random, 54% capability-only). The only coefficient that survives both cluster-robust inference and Holm–Bonferroni is the single-agent baseline (Probust = 0.004, PHolm = 0.018). A separate decision rule puts a capability-saturation threshold near 45%: above that, expected MAS gain is zero or negative. The rule matches the sign of the MAS delta in 94% of 16 model × benchmark cells on SWE-bench Verified and Terminal-Bench. Trace-level error amplification ranks architectures cleanly: SAS 1.0×, centralized 4.4×, hybrid 5.1×, decentralized 7.8×, independent 17.2×. Realized overhead versus SAS: independent 58%, decentralized 263%, centralized 285%, hybrid 515%. Hybrid uses 6.2× as many reasoning turns (44.3 vs 7.2).
Leave-one-dataset-out R² is −2.09. Absolute success does not travel across domains. Relative ranking of architectures still does, inside the tested set.
For anyone shipping agents, the threshold is more useful than the slogan. If SAS already clears 45%, skip the orchestrator. Parallel information gathering (finance-style) is where centralized teams pay rent. Tool-heavy work with mutual checks leans decentralized. Sequential planning and high-baseline bugfixing stay single-agent. The authors call 45% a practical selection rule, not a universal scaling law. With only six benchmark clusters, tool-count × efficiency interactions fail cluster-robust tests and stay directional.
This is the tool-using cousin of budget-matched multi-hop studies that also find SAS on top once thinking tokens are equalized. Here the hinge is decomposability, not just token accounting.
SWE-bench Verified and Terminal-Bench use 20-instance subsets; bootstrap intervals are about ±20 points, so cell-level pairwise tests are weak. The 94% figure is an aggregate sign match across 16 cells. Prompts were not tuned per model family. In 13 heterogeneous mixing trials, centralized mixed teams lag strong-model homogeneous teams by 12.6 points and do not beat the saturation threshold. All communication is natural-language tokens; whether latent sharing would rewrite the curve is untested. Cluster-robust SEs with G = 6 are conservative, and several naive-OLS effects were demoted to description. No embodied, multi-user, or long-horizon feedback settings.