Don't buy multi-agent. Don't build it either.

You know the slide. Five rounded boxes: Planner, Researcher, Writer, Critic, Manager. Arrows between them. A sentence about agents "collaborating like a high-performing team". Somewhere near the bottom, a price.

The boxes are real in the sense that someone drew them. This essay is about what is underneath, which is much less than the slide suggests and which you can have today without buying or building anything called multi-agent.

01The question that ends the demo

Point at any arrow on the slide and ask: who decides that this model call happens next? Only two answers exist. A line of code decides, or a model's output does.

notes = research(task)
draft = write(notes)
final = critique(draft)

tools = [search, edit,
         spawn(prompt)]
loop(task, tools)

The first answer has had a plain name since 2024, when Anthropic defined workflows as models and tools "orchestrated through predefined code paths" and kept the word agent for systems where the model directs itself.¹ A workflow is a fine thing to build. It is also a Python file.

The second answer is one agent, because a tool call is the only way a model decides anything. The vendors say so in their own documentation. OpenAI: "Handoffs are represented as tools to the LLM."² Claude Code starts a subagent through a tool named Agent, gives it its own context window and hands a summary back to the caller.³ The lead agent in Anthropic's Research feature creates its parallel searchers by calling tools.⁴ Cognition's manager Devin reaches its child Devins through an internal MCP server, meaning tools again.⁵

No model in a multi-agent system can tell that it is in one. The callee sees a prompt. The caller sees a tool result. The team exists on the slide.

02The brochure, translated

The last row is the one with numbers attached.

03The receipts

Start with the success story, because it is the honest one. Anthropic's lead-plus-subagents system beat a single Claude Opus 4 agent by 90.2% on an internal research eval. The same post explains why: token usage alone accounted for 80% of the performance variance on BrowseComp, and the authors write that such systems work mainly because they "help spend enough tokens to solve the problem".⁴ The architecture was a way to spend a big budget in parallel. That is useful. It is also a spending pattern, and you can have a spending pattern without a framework.

Now control the experiment. A Google Research team ran 180 configurations across five architectures, four benchmarks and three model families, with tools and prompts standardised.⁶

Four bars you cannot tell apart, next to a column of red. The paper's mean multi-agent effect is −3.5%, with a 95% interval from −18.6% to +25.7%. The same team that lifts a parallelisable finance task by 80.8% sinks a sequential planning task by 50.3%. A product that helps or hurts by half depending on the task shape should be sold with that sentence on the box.

Three 2026 papers close the remaining exits. Tran and Kiela hold reasoning tokens equal and find single agents match or beat multi-agent setups on multi-hop reasoning, across three model families.⁷ Xu and colleagues replay published multi-agent workflows as one agent in one multi-turn conversation: across seven benchmarks it reaches the same performance and costs less, because the cache is reused.⁸ Ao, Gao and Simchi-Levi prove the general case: a delegated network that receives no new outside signal cannot beat one decision maker holding the same information.⁹

As for the failures: in Berkeley's MAST study, about a third of them were inter-agent misalignment, a category that has no equivalent inside a single context.¹⁰ The framework creates the coordination problem and then offers features for managing it.

04Even the believers folded

In June 2025 Cognition's Walden Yan published "Don't Build Multi-Agents".¹¹ Ten months later he described what Cognition now ships under that name: a code reviewer that starts with no shared context, a small model that calls a bigger one for hard problems, and a manager that delegates to children, all on the condition that it works "when writes stay single-threaded".⁵ Each of those is one loop calling another through a tool.

Anthropic wrote the most cited multi-agent success story. On 22 October 2026 its Applied AI team told enterprise builders to fold subagents back into a single loop, one at a time, keeping only the ones whose removal makes the evals worse. Where a subagent survives, the advice is to call it as a tool. One customer's single loop ran about twice as fast as its multi-agent harness.¹²

The skeptic and the pioneer now give the same instruction. Neither one mentions a framework.

05What is real, so nobody says I hid it

Strip the word and one real thing is left: a call that runs in a fresh context window. It is worth making in exactly three situations, and all three are in your existing agent runtime.

1. The context is rotting. Long tasks degrade a single window. A reset at a measured depth helps: in an October 2026 preprint, a decomposed system won a ledger task at 20, 50 and 100 steps, close to where a simple cost model predicted.¹³
2. The work is parallel and read-only. Ten searches at once finish sooner than ten in a row. One writer at the end.
3. You want a second opinion that is second. A reviewer that has not seen the author's reasoning catches what the author rationalised. Cognition's clean-context reviewer finds an average of two bugs per pull request, about 58% of them severe.⁵

There is also one case that earns the name. When agents answer to different owners, a buyer's and a seller's, each holds information the other cannot be given and no single context may hold both. That is a multi-agent system in the sense economists have used for decades. If your vendor is selling you that, they are selling you a protocol between companies, and the conversation is a different one.

Inside one team with one goal, there are no agents to count. There are places where you cut the context, and each cut should have to justify itself.

06Five questions for the next pitch

1. Who decides the next call, your code or the model?Either answer describes something you already have.
2. What does it score against one agent given the same tokens?If nobody measured, the benchmark compared budgets.
3. What crosses between contexts, and what gets dropped?The summary is where the information dies.
4. How many writers?The right number is one.
5. What does each agent know that the others cannot be told?If the answer is nothing, the theorem says the team adds nothing.

Then run the test on your own architecture document. Replace every "agent" with "call". The sentences that still make a decision are your design. The rest was the slide.

References

1. Anthropic, Building effective agents, 19 December 2024.
2. OpenAI Agents SDK documentation, Handoffs and Orchestrating multiple agents.
3. Claude Code documentation, Subagents.
4. Anthropic, How we built our multi-agent research system, June 2025.
5. Walden Yan, Multi-Agents: What's Actually Working, Cognition, 22 April 2026.
6. Yubin Kim et al., Towards a Science of Scaling Agent Systems, arXiv 2512.08296, December 2025; Google Research summary, 28 January 2026.
7. Dat Tran and Douwe Kiela, Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets, arXiv 2604.02460, April 2026.
8. Jiawei Xu et al., Rethinking the Value of Multi-Agent Workflow: A Strong Single Agent Baseline, arXiv 2601.12307, January 2026. The result covers workflows whose agents share one base model.
9. Ruicheng Ao, Siyang Gao and David Simchi-Levi, On the Reliability Limits of LLM-Based Multi-Agent Planning, arXiv 2603.26993, March 2026.
10. Mert Cemri, Melissa Z. Pan, Shuyi Yang et al., Why Do Multi-Agent LLM Systems Fail?, NeurIPS 2025 Datasets and Benchmarks Track, arXiv 2503.13657v3. Failure rates range from 41% to 86.7%; the inter-agent share (about 32%) is the sum of the paper's per-mode percentages.
11. Walden Yan, Don't Build Multi-Agents, Cognition, 12 June 2025.
12. Diane Sarkis, Building agents: When to fold subagents into your main agent loop, Anthropic, 22 October 2026.
13. Akshit Anchan and Nayonika Sen, Attention Tax, Handoff Tax: A Stylised Model of When Multi-Agent LLM Systems Help, arXiv 2610.06069, October 2026. Preprint.