Key takeaways
- In a clinical prediction study, the best multi-agent method reached an F1 of 87.50 at 15.5K tokens per case against a single agent using roughly 1.9K. That is about eight times the spend.
- In an activity recognition study, multi-agent collaboration used 4x the tokens of a 15,000-token single-agent baseline and moved accuracy from 50.5% to 79.5%.
- The direction of the trade is consistent across papers: more agents buy accuracy, and they buy it with tokens. Nothing published suggests coordination is free.
- Multipliers of 515% and ranges of 4x to 220x circulate widely in practitioner write-ups. We could not trace either to a primary paper and do not repeat them as findings.
- The decision is not architectural taste. It is whether the accuracy your task needs sits above what one agent delivers, at a price you are willing to pay per completed task.
Does ChatGPT recommend your competitor instead of you?
Check how AI assistants describe your company, and whose name they give when someone asks for a recommendation in your category. Run an audit today, from $19.
Two agent papers published this year put a price on coordination. In a clinical prediction study, the strongest multi-agent method reached an F1 of 87.50 while consuming 15.5K tokens per case. The single-agent comparison in the same paper used roughly 1.9K.
That is about eight times the spend for a result the authors judged worth reporting. Whether it is worth it in your fleet depends on a number most teams do not calculate.
How much does coordination actually cost?
The honest answer is that it depends on the shape of the work, and the published range is narrower than the numbers circulating on the open web suggest.
| Study | Single agent | Multi-agent | Multiplier | Accuracy effect |
|---|---|---|---|---|
| Clinical prediction, arXiv:2604.00085 | ~1.9K tokens | 15.5K tokens | ~8x | F1 87.50, highest in the comparison |
| Activity recognition, arXiv:2605.04608 | 15,000 tokens | ~60,000 tokens | 4x | 50.5% to 79.5% |
Two data points are not a distribution. They are enough to establish the direction, which is that coordination is paid for in tokens and the bill is measured in multiples rather than percentages.
About the numbers you will see elsewhere
Search for this topic and you will meet a figure of 515% additional tokens, and a range of 4x to 220x attributed to a university study. Both appear across multiple 2026 write-ups. We could not trace either to a primary paper, so they do not appear as findings here. A number you cannot follow to its source is not evidence, however often it is repeated.
Why does coordination make some tasks worse?
Because passing work between agents costs more than tokens. It costs fidelity.
Every handoff serialises one agent’s understanding into text and asks the next to reconstruct it. On a task that decomposes cleanly, that is a small loss against a large parallelism gain. On a task where step four depends on a nuance established in step two, the reconstruction is where the answer quietly changes.
This is why sequential reasoning is the category where multi-agent designs disappoint most often. The work does not decompose, so every boundary you introduce is a lossy copy with no compensating benefit.
Which problems genuinely need more than one agent?
Three, in our reading of the literature and of how teams describe their production systems.
- Genuinely parallel sub-tasks with independent context. Six documents to summarise is six agents. One document to reason about across six steps is one agent.
- Different tools or different models per sub-task. If one step needs a code sandbox and another needs a retrieval index, separating them buys you smaller tool definitions and cheaper models where the work is easy.
- Explicit boundaries for audit. Sometimes the reason to separate is not performance. It is that you need a defensible record of which component decided what.
Notice that none of these is “the task is hard”. Difficulty is an argument for a better model, not for more agents.
How do you test whether your design earns its cost?
Price the task, not the call. This is the calculation the papers make and most fleets do not.
| Step | What to record | Why it matters |
|---|---|---|
| 1. Fix the input set | The same 200 real tasks for both designs | Different inputs make the comparison meaningless |
| 2. Define acceptance before running | The check that decides pass or fail | Deciding afterwards is how you talk yourself into the answer you wanted |
| 3. Count total tokens, not per call | Every call in the run, including retries | Coordination overhead hides in the calls nobody counts |
| 4. Divide by tasks that passed | Cost per completed task, both designs | The only figure that survives contact with a budget |
The fourth row is the one that changes minds. A design that is 8x the token cost and 30 points more accurate can easily be cheaper per completed task than a single agent you have to run three times and check by hand.
The question to ask before adding the second agent
Not “would another agent help”. Almost always yes, marginally. Ask instead: what is my cost per completed task today, and what would it have to become for this change to be a mistake? If you cannot answer the first half, the architecture is not your problem yet.
Frequently asked questions
How many more tokens does a multi-agent system use than a single agent?
Published comparisons put it between four and roughly eight times for the configurations they tested. A clinical prediction study reported a single agent at about 1.9K tokens per case and the strongest multi-agent method at 15.5K. An activity recognition study reported 4x against a 15,000-token single-agent baseline.
Is a multi-agent system more accurate than a single agent?
In the studies above, yes, and by large margins on the tasks they chose. Activity recognition accuracy moved from 50.5% to 79.5%. That is a task-specific result, not a general law, and the papers report it alongside the token cost rather than instead of it.
When should you use a single agent instead of multiple agents?
When the work is a linear pipeline, when it fits comfortably in one context window, and when no sub-task needs a different tool set or a different model. A single agent is cheaper, faster to debug and produces a trace you can read in one pass.
How do you measure whether multi-agent coordination is worth the cost?
Price the task, not the call. Divide total token spend by completed tasks that passed your acceptance check, for both designs, on the same inputs. If the multi-agent version does not clear the single-agent baseline on that number, the extra agents are decoration.
Method and sources
- Elias Lumer and colleagues, Don't Break the Cache: An Evaluation of Prompt Caching for Long-Horizon Agentic Tasks, arXiv:2601.06007, 3 February 2026. Consulted for how repeated context behaves across agent turns.
- One Panel Does Not Fit All: Case-Adaptive Multi-Agent Deliberation for Clinical Prediction, arXiv:2604.00085. Source for the F1 of 87.50 at 15.5K tokens and the single-agent figure of roughly 1.9K tokens.
- SensingAgents: A Multi-Agent Collaborative Framework for Robust IMU Activity Recognition, arXiv:2605.04608. Source for the 4x token figure against a 15,000-token single-agent baseline and the accuracy move from 50.5% to 79.5%.
- Token Economics for LLM Agents: A Dual-View Study from Computing and Economics, arXiv:2605.09104. Consulted for framing of per-task cost accounting.
- Figures of 515% additional tokens and a 4x to 220x range appear in several 2026 practitioner write-ups. We were unable to trace either to a primary study and have not used them as findings in this article.