More agents do not always give better results. The fit between the agent architecture and the task structure decides the result.
Each figure has two buttons. Copy DataviewJS copies a block that runs the interactive figure in Obsidian (Dataview plugin). Copy HTML copies a static snapshot for any Markdown note. Each formula has a Copy LaTeX button.
The authors compare one agent with four types of agent teams under the same tools, prompts and token budget. §1, §4.1
Teams help a lot on tasks that divide into parallel parts, for example financial analysis (+80.8%). Teams hurt a lot on tasks with strict step order, for example Minecraft planning (−70.0%). §4.2, Fig. 2
If one agent already solves about 45% of a task or more, extra agents usually add cost but not accuracy. This "capability saturation" effect is the most robust result of the paper. §4.3, §4.5, §6
The five designs differ only in who sends messages to whom. Each example has 3 agents. Moving dots are messages. Solid arrows send work or messages; dashed arrows return results. §3.1, Table 2
Select an architecture. The diagram shows how messages move between agents. The chart shows what the architecture costs (turns) and what it gives (success rate). The strips below compare all six metrics of Table 5. All values are averages across the six benchmarks. §3.1, Table 2, Table 5
The chart shows the main point. All five architectures stay in a narrow band of success rate (0.370–0.477). But the number of turns goes from 7.2 to 44.3. Thus, coordination uses the budget, and it does not by itself buy accuracy. Table 5
An agent is an LLM that runs a loop: it reasons, calls a tool, reads the result and then acts again. §1, §3.1 A single-agent system (SAS) has one such loop. A multi-agent system (MAS) has two or more agents that send messages to each other. §3.1
Earlier studies compared architectures with different prompts, tools or budgets. Thus, nobody could say if a gain came from the architecture or from the setup. §1 This paper keeps the prompts, the tools and the total token budget the same. Only the architecture and the model change. §1, §4.1
The six benchmarks are all agentic. An agentic task needs many steps of interaction with an environment, and some information is hidden until the agent asks for it. §3.2, Table 1
The nine models come from OpenAI (GPT-5 nano, mini, full), Google (Gemini 2.0 Flash, 2.5 Flash, 2.5 Pro) and Anthropic (Claude Sonnet 3.7, 4, 4.5). Their Intelligence Index goes from 42 to 71. §4.1, App. A
Each cell shows the relative change of one MAS architecture against the SAS on the same benchmark. Green means the team did better. Orange means the team did worse. Point to a cell to read the details. §4.2, Fig. 2
Each cell above uses this relative change Δ. S̄ is the mean success rate. Fig. 2
The mean change across all benchmarks and architectures is −0.3%. The 95% interval goes from −58.7% to +77.2%. Thus, "MAS is better" and "MAS is worse" are both wrong as general statements. §4.2
Domain complexity alone does not explain the pattern. Finance-Agent (D = 0.41) and PlanCraft (D = 0.42) have almost the same complexity score, but their results are opposite. The difference is decomposability: Finance divides into parallel parts, and PlanCraft does not. §4.2, Table 11
The paper shows short execution traces for the two extreme cases. §4.2
The authors fit one regression model to all 260 configurations. The model predicts performance from model capability, task properties and measured coordination metrics. It has no parameters for specific datasets. §4.3, Eq. 1
The interaction of the single-agent baseline PSA with agent count has a negative coefficient (β = −0.236, p = 0.004). Thus, when one agent already does well, more agents give less. Table 4 The decision boundary is β₄ / β₁₇ ≈ 0.040 / 0.236 ≈ 0.17 in standardized units. That is about 45% in raw performance. §4.3
Under cluster-robust errors, this interaction becomes a directional pattern (p = 0.105). But the main effect of PSA stays significant (p = 0.004). For this reason, the authors call capability saturation the most robust finding. §4.5, Table 14
The SWE-bench Verified result agrees with this rule. Most models already score above 45% alone, and all four MAS types score lower than SAS (−2.1% to −14.9%). §4.2 On SWE-bench and Terminal-Bench, the threshold matches 94% of the 16 model–benchmark cases (binomial test, p < 0.001). §4.5
The interaction of efficiency with tool count is negative (β = −0.096, 95% CI [−0.154, −0.037], p = 0.002). Tool-heavy tasks lose more from the inefficiency of teams. §4.3, Table 4 This term survives the Holm–Bonferroni correction (p = 0.026), but not cluster-robust errors (p = 0.205). Thus, the paper reports it as a directional pattern. §4.5, Tables 14–15 The mechanism: a team divides the fixed token budget across agents. Then each agent has fewer tokens for complex tool use. §1 On simple tasks with 4 tools or fewer, the effect is very small (|ΔP| < 0.05). §4.3
Trace-level error amplification measures how much extra work coordination failures cause, against the SAS (= 1.0). Independent agents have no check step, so errors go directly to the output (17.2×). An orchestrator checks the work before it combines it (Centralized, 4.4×). §1, §4.4, Table 5
Read this pattern with care. In the regression, error amplification is not significant after the model controls for the other metrics (β = 0.014, p = 0.658). The authors say that efficiency and overhead explain the differences better. §4.3
The authors sort errors into four categories. Each architecture changes the rates in a different way. §4.4
| Error type | SAS baseline | Centralized | Decentralized | Hybrid | Independent |
|---|---|---|---|---|---|
| Logical contradiction | 12.3–18.7% | 9.1% | 11.5% | n/s | 16.8% |
| Numerical drift | 20.9–24.1% | 18.3% | 18.3% | 26.4% | 23.2% |
| Context omission | 15.8–25.2% | 8.3% | 11.2% | n/s | 24.1% |
| Coordination failure | — | 1.8% | 3.2% | 12.4% | 0% |
Coordination failure occurs only in MAS. Hybrid has the highest rate because its protocol is the most complex. §4.4
The total number of turns grows faster than the number of agents. The fit is a power law over the architecture means. §4.4
Fit quality: R² = 0.974. The 95% interval of the exponent is [1.685, 1.763]. §4.4
The token budget stays fixed. Thus, each agent gets less space to reason as the team grows. The authors conclude that per-agent reasoning becomes too thin beyond 3–4 agents. §4.4 The paper tested up to nine agents only, so values above 9 on the slider are extrapolations. §5 (i), Fig. 5
The paper gives three task archetypes as examples of how to use the model for architecture selection. §4.3
The baseline is above the threshold, and the tool count is low.
It controls errors (4.4×) and keeps the overhead manageable.
Parallel work and redundancy outweigh the efficiency loss, although overhead is high (263%).
The third archetype has a baseline above 45% but still favors a team. Thus, the 45% rule is a strong trend, not an absolute law. §4.3 On held-out configurations, the model selects the best architecture in 87% of cases. Random choice gives 20%, and a model that uses capability only gives 54%. §4.3
This table shows all 19 coefficients of Eq. 1 with their 95% intervals, exactly as Table 4 gives them. All predictors are standardized after transformation, and intelligence is centered at Ī = 57.5. The intercept is 0.430 [0.412, 0.448]. Table 4
Table 4 reports naive OLS p-values. Two stricter checks give different results for some terms. The last two columns show these values where the paper reports them. §4.5, Table 14, Table 15
The full model reaches R² = 0.463 on training data and R² = 0.373 in five-fold cross-validation. If the authors replace the Intelligence Index with the Agentic Capability Index (ACI), the cross-validated R² goes up to 0.413. §4.3, Table 3, Table 13 None of the five coordination metrics is significant as a main effect. Only some of their interactions are significant: efficiency × tools, error × baseline and redundancy × agents. Table 4
The authors run three robustness checks. The results make most findings weaker, but one finding stays strong. §4.5
Other limits, as the authors state them:
The paper has 67 references in 13 groups. Open the full interactive citation map. It has a reader panel and a role for each reference. reference list, §1–§4
Writing style: STE-lite, based on ASD-STE100 Simplified Technical English.