Amazon researchers compared one agent with several coordinated agents on 13 enterprise workflows.
Multi-agent setups were more accurate, but used 1.4 to 4.7 times as many tokens.
Their conclusion: make a single agent the default, and add more only where accuracy pays for it.
Should a business workflow run on one AI agent, or on a team of agents that hand work to each other? A study from Amazon researchers, posted to arXiv in early November 2026, put numbers on the trade-off.
What they tested
The researchers ran single-agent and multi-agent versions of 13 enterprise workflows and measured two things: how often the job came out right, and how many tokens it took to get there.
What they found
- Accuracy: the multi-agent versions scored higher.
- Cost: the more coordinated setups used between 1.4 and 4.7 times as many tokens as the single agent doing the same job.
Agents that talk to each other spend tokens on the conversation itself: planning, handing over context, checking each other's work. That overhead is paid on every run.
How to use this
- Start with one agent. Build the workflow as a single agent with good tools, and measure its accuracy on real cases.
- Add agents only where the misses are expensive. If a wrong answer costs more than 1.4–4.7 times the tokens, a second agent that checks the first may pay for itself.
- Use cheaper models for the extra agents. A reviewer or router agent rarely needs the same model as the one doing the work. On WayJet you can give each agent a different
modelbehind the same API key, and see each one's spend on the Usage page.
The study's headline is simple enough to act on: one agent is the sensible default, and every extra agent should earn its tokens.
Источники и контекст
- Crypto Briefing: Amazon study finds a single AI agent is the best default for enterprise workflowscryptobriefing.com · Проверено 12 нояб. 2026 г.

