A paper by researchers at Amazon and the University of Washington finds that the one multi-agent design that significantly beat a single AI agent on enterprise workflows used 4.7 times the tokens per task to get there, lifting task success to 75.6% from 67.4%, an 8.2-point gain. The preprint, Agent Architecture Cost-Success Trade-offs, was submitted to arXiv on 7 November.
By our arithmetic from the paper’s Table 3, that is about 155,200 tokens per successfully completed task against about 36,900 for the single agent, 4.2 times as many. The authors call the single agent “the efficient default”.
Business Pill · Multi-Agent Systems
A one-minute explainer of multi-agent systems: how several AI agents split and share work. It teaches the general idea only and does not discuss this paper.
The key insight: As we read it, the paper turns the agent-architecture choice into a cost question: on these 13 procedures the single agent set the price floor, and the one design that cleared it significantly paid 4.7 times the tokens per task for 8.2 points of task success.
The team held the model (Claude Opus 4.8), the tasks and the tools fixed and varied only the architecture. They compared four designs, a function-calling single agent, a skill-augmented agent, centralized multi-agent orchestration and a triage swarm, run as eight arms across 13 enterprise standard operating procedures (SOPs) from the SOP-Bench benchmark, among them know-your-business checks, patient intake, dangerous goods and warehouse tasks.
The paper reports that every coordinated arm, multi-agent or skill-augmented, “spends 1.4โ4.7ร the tokens and 2.6โ5.8ร the latency of the single agent for gains that are usually small and rarely significant”. In its words, the single-agent baseline is “cheapest on every axis”, yet competitive on accuracy, with no arm both cheaper and more accurate.
Tokens per task for each of the eight architecture arms in the Amazon paper, with each arm’s task success rate. Orchestration v2, the only arm to significantly beat the single agent, used 117,300 tokens per task against 24,900. Source: arXiv 2611.12953, Table 3.
The 8.2-Point Exception
One design did separate. The authors’ first orchestration build scored 58.2%, the lowest of any arm, because each worker saw only its slice of the procedure. A redesign that gave every worker the full SOP and tool set plus a re-dispatch round, orchestration-v2, reached 75.6%, the only arm to significantly beat the 67.4% baseline. It also cost the most: 117,300 tokens per task against 24,900, and 3.1 times the latency.
The authors read this “not as” a win for multi-agent systems but as “implementation quality mattering more than the single- versus multi-agent choice”. High-quality skills lifted the skill-augmented agent to 73.1%; a degraded skill set dropped it to 59.8%, below the single agent.
Task success for each arm against the 67.4% single-agent baseline, with each arm’s token multiple from Table 3. Orchestration v2 (75.6%, 4.7x tokens) is the only arm the paper finds significantly better. Source: arXiv 2611.12953.
How This Differs From the Agent-Teams Result
Last week we covered a Duke, Modal Labs, Princeton and NYU preprint in which ten agent teams of ten scored 27% above 100 solo agents. That study spends a fixed budget of agents on tasks a verifier can score and keeps the best result; the Amazon paper holds one model and measures what each extra layer of coordination costs on business procedures. The two answer different questions, and the Amazon authors say architecture matters most on complex, branching SOPs, a trend they call directional and not yet significant.
Their advice to practitioners is to “start simple, measure marginal TSR on the target SOPs”, fix implementation failures first, and switch architecture only when the extra success justifies the extra cost. “Architectural complexity should be earned by evidence, not assumed to be an upgrade”, they write.
The Structural Read
As we read the orchestration result, the same multi-agent family went from the worst arm (58.2%) to the best (75.6%) once each worker saw the full procedure, which is why the authors say implementation quality matters more than the single- versus multi-agent choice.
As we read the contrast with the agent-teams preprint we covered, teams pay off when a verifier can pick the best of many attempts; on business procedures with one answer per task, the Amazon authors measure the extra coordination mostly as cost.
Nandi, Arora et al., Amazon and University of Washington, arXiv 2611.12953
“Architectural complexity should be earned by evidence, not assumed to be an upgrade”
Three Implications
FOR TEAMS BUILDING AGENTS The paper reports every coordinated arm spent 1.4 to 4.7 times the single agent’s tokens; the authors advise starting simple and measuring marginal task success first.
FOR FINANCE OWNERS OF AI BUDGETS By our arithmetic the top arm cost about 155,200 tokens per successful task against about 36,900, 4.2 times as many.
FOR VENDORS OF MULTI-AGENT TOOLING The authors found added structure only plausibly pays off on complex, branching procedures, a trend they call directional and not yet significant.
The framework argues: “The surface, not the model, governs autonomy.”
As we read it, the Amazon paper holds the model fixed and changes only the surface around it, and task success moves from 58.2% to 75.6%. The framework predates this paper and does not discuss it.
What Is Not Established
The authors say all results come from one benchmark, SOP-Bench, whose synthetic mock tools omit the latency, errors and nondeterminism of live services, and every arm ran on one model, so they make no claim that the ranking or cost multipliers generalize to other model families. Each architecture is a single implementation, and most differences between arms are not statistically significant.
The paper is a preprint and has not been peer reviewed, and the tokens-per-successful-task figures are our arithmetic, not the authors’. It does not say whether Amazon uses these findings in its own products. We did not contact the authors or Amazon.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Amazon’s researchers found that on 13 enterprise workflows a single AI agent was the cheapest design and competitive on accuracy; the only multi-agent build that significantly beat it gained 8.2 points while spending 4.7 times the tokens per task, about 4.2 times per successful task by our arithmetic.
A note on sourcing. We read the full preprint, arXiv 2611.12953, submitted on 7 November 2026 by researchers at Amazon and the University of Washington. It has not been peer reviewed. Tokens-per-successful-task figures are our arithmetic from its Table 3. We first saw the paper in a post on X by @rohanpaul_ai. We did not contact the authors or Amazon.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
