Amazon official logo (public domain, Wikimedia Commons) — CryptoBriefing brand treatment
Amazon researchers found that multi-agent setups burned up to 4.7 times more tokens for modest and often statistically insignificant gains
The enterprise AI world has spent a lot of energy building teams of AI agents that hand work to one another. A new Amazon study suggests most companies should start with one.
Amazon researchers found that a single, well-built agent is usually the best default for enterprise workflows. The more complex setups consumed far more tokens and time, and in many cases delivered only small improvements in exchange.
What the researchers tested #
The paper is titled “Agent Architecture Cost–Success Trade-offs: A Controlled Evaluation across Enterprise-Style SOPs” and was posted to arXiv as arXiv:2611.12953 in early November 2026. SOPs, or standard operating procedures, are the step-by-step playbooks companies use for routine work.
The team ran agents through 13 realistic enterprise workflows. They compared four approaches.
The first was a function-calling single agent. That is one model that can call external tools, such as looking up a record or submitting a form.
The second was a skill-augmented agent. This is the same solo agent, equipped with prepackaged skills that guide how it handles specific tasks.
The third was centralized multi-agent orchestration. One coordinating agent splits up the work and delegates pieces to specialist agents.
The fourth was a triage swarm. Here a group of agents routes tasks among themselves based on what each one handles.
The main yardstick was task success rate, or TSR: the share of tasks the system completed correctly.
The numbers behind the verdict #
The single-agent baseline hit a TSR of 67.4%. That is the number everything else had to beat.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Adding high-quality skills lifted success to 73.1%. Adding low-quality skills dragged it down to 59.8%, which is worse than doing nothing at all.
The multi-agent results were all over the map. The weakest centralized multi-agent configuration scored 58.2%, well below the solo baseline.
Meanwhile, one redesigned multi-agent variant reached 75.6%. That was the top score in the study and a clear step above the baseline.
The more coordinated setups consumed between 1.4 and 4.7 times as many tokens as the single agent. Latency rose by 2.6 to 5.8 times compared to the single-agent arm.
Tokens are the units of text that AI models process and that providers bill for. More tokens means a bigger invoice, and higher latency means users wait longer for answers.
No universal winner #
Perhaps the most important finding is what the researchers did not find. No architecture delivered a statistically significant improvement across all 13 SOPs.
Many of the gains from skills were modest. Some did not clear the bar for statistical significance either.
Performance depended heavily on the kind of workflow involved. The researchers saw directional benefits from added structure on more complex tasks, though those improvements had not reached statistical significance.
The study also points to implementation quality as a bigger factor than architecture. The gap between the best multi-agent variant at 75.6% and the worst centralized one at 58.2% came from how the systems were built, not just which category they belonged to.
Why this matters for companies building AI agents #
The core recommendation is to treat the single agent as the starting point and add complexity only when a specific workflow shows a clear, measurable payoff.
Teams evaluating agent designs would need to weigh any success-rate gain against token consumption that could run up to 4.7 times higher and response times that could stretch up to 5.8 times longer.
If low-quality skills can push success from 67.4% down to 59.8%, then curating, testing and maintaining those skills becomes real engineering work rather than an afterthought. A poorly built orchestration layer can underperform the simple baseline, as the 58.2% result shows, while a carefully redesigned one can lead the field.
There are limits worth keeping in mind. The study covered 13 workflows, and the hints that complex tasks benefit from added structure were not statistically significant.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.