When multi-agent is worth 15 times the tokens

Multi-agent systems beat single-agent systems on Anthropic's own research evals. They also burn around fifteen times the tokens getting there. Both facts are true, and most of the advice I read treats only the first one as if it settles the question.

I route every research request through a three-tier decision before I let anything spawn a fleet, and the tiering exists specifically because the fifteen-times number is real money and real latency, not an abstraction. Tier zero is a lookup: one fact, a handful of sources, the kind of thing that fits in a reply and doesn't need a durable record. That gets an inline search, no agent at all. Tier one is standard: one topic, five to eight sources, low risk that the sources actively contradict each other. That's a single research agent producing one structured file. Tier two, the fifteen-times tier, is reserved for genuinely broad questions with many sources that plausibly disagree, where the answer needs adversarial verification before it's trustworthy.

The default, when I'm not sure which tier a request belongs in, is tier one, not tier two. That's a deliberate bias. Multi-agent's advantage on Anthropic's evals doesn't transfer automatically to every question that looks broad; it transfers to questions where the breadth is real, where sources actually conflict, and where a wrong synthesis would cause real downstream damage if nobody checked it. A single well-scoped agent handles most requests that merely look complicated on the surface, at a fraction of the cost, and its output is cheap to supersede if it turns out to be wrong.

What made this concrete for me was going and finding the actual controlled study behind the "decompose into agents" advice that gets repeated constantly with almost no citation attached. A DeepMind paper running two hundred and sixty configurations found that a regression model using three measurable properties of a task, how decomposable it is, how many tools it needs, how deep its sequential dependencies run, predicted the best architecture with eighty-seven percent accuracy on held-out tasks. A capability-only model, ignoring task structure entirely, only hit fifty-four percent. In other words: whether multi-agent helps is mostly a fact about the task, not a fact about how good your agents are. The task property that matters most is whether it's genuinely parallelisable and whether coordination between agents is centralised rather than freeform.

The token overhead numbers from that same study are what made the fifteen-times figure feel concrete rather than abstract. Independent multi-agent architectures ran fifty-eight percent over a single-agent baseline. Decentralised coordination, two hundred sixty-three percent. Centralised, two hundred eighty-five percent. Hybrid, the worst of the set, five hundred fifteen percent, specifically on tool-heavy tasks where coordination overhead compounds instead of adding. That's a range that depends entirely on how the coordination is structured, not a fixed fifteen-times tax across the board, and the worst-structured version costs more than five times a single agent for work a single agent might have done just fine.

There's also a ceiling I didn't expect going in: past roughly forty-five percent single-agent-baseline accuracy on a task, adding more agents turns net negative. If your one well-scoped agent is already doing reasonably at something, fragmenting it into a fleet is statistically more likely to make things worse, not better. This runs directly against the "just add more agents" instinct that a lot of 2026 production advice leans on, some of it citing case studies that, when I went and checked the primary source directly, turned out to have no methodology, no baseline, and no named data behind the headline claim at all.

So the actual test I apply before spawning a fleet for anything, research or otherwise, is three questions, and all three have to clear. Does this genuinely decompose into independent, parallelisable work, not a sequential chain dressed up as parallel. Will coordination be centralised with a strict, schema-disciplined handoff between agents, rather than agents passing raw context back and forth. And is a single well-scoped agent already failing at this, not just looking complicated. If any of those three fails, I chain the work through one continuous context instead, and I save the fifteen-times spend for the narrow slice of work where sources genuinely conflict and getting it wrong would actually cost something. Most work isn't that. Most work just feels that way because it has a lot of moving parts, and moving parts aren't the same thing as genuine independence.