The burden of proof is on splitting

Multi-agent systems look like org charts and cost like them. Why one excellent prompt is the right default, what a split really costs, and the one split that is almost always worth paying for.

Andriy Kudelya · 2026-07-11 · 8 min

Hand someone a hard job and a box of AI agents, and the first diagram they draw is an org chart: a researcher, a writer, an editor, a reviewer, maybe a project manager to keep them honest. It feels rigorous. It looks like how serious work gets done. We drew that diagram too — and then we started counting what it costs and checking what it actually buys.

The instinct comes from a category error. We split work across people because of human constraints: one head can't hold the whole job, attention is serial, expertise is narrow. Splitting buys parallel attention and pays for it in coordination — meetings, handoffs, misunderstandings. Language models invert both sides of that trade. A single context window can hold more of the job than any human working memory ever could. And coordination between model contexts is dramatically worse than between colleagues: an agent can't lean over and ask, there is no hallway, no shared history — every handoff is a lossy serialization of one context into text for another.

Split a job across agents and you pay the coordination tax without the constraint that justified it.

What a split actually costs

Tokens, first and plainly. Anthropic, describing their own production research system, report that a single agent uses about 4× the tokens of a chat interaction, and multi-agent systems about 15×. That's not an argument against multi-agent — they shipped one — it is a price tag. A 15× premium needs a job whose value clears it, and most jobs don't.

Handoff loss. When one agent passes work to another, an artifact travels; the understanding doesn't. Every judgment call the first agent made silently — how it framed the problem, which interpretation of an ambiguous term it chose, what it decided not to include — is invisible in the artifact and unavailable to the next context. Each boundary is a compression, and the losses compound down the chain.

Coupling. This is the failure mode Cognition's “Don't Build Multi-Agents” essay nailed: actions carry implicit decisions. Give two parallel workers halves of one design task and each will resolve the shared ambiguities its own way — different naming, different framing, different unstated assumptions — and the merged whole is incoherent even though every part is locally fine. Parallel agents don't disagree loudly; they disagree silently, in the artifact.

Debugging surface. One prompt has one place to be wrong. A five-agent system has five prompts, four handoffs, and an orchestration policy — and when the output is bad, the failure is distributed across all of them. You haven't decomposed the problem; you've decomposed the blame.

The four criteria that justify a split

A split is an engineering decision, and like any engineering decision it should be able to name the force that requires it. We recognize exactly four.

1. Fresh eyes are the point. A reviewer that shares the author's context inherits the author's blind spots — it will approve the same mistake for the same reason the mistake was made. Verification is the one place where not knowing is the asset: a gate that sees only the artifact and a rubric, and must find evidence rather than recall intentions. This is the split that is almost always worth paying for, and note why it works: it is valuable precisely because it refuses the shared context that makes every other split expensive.
2. True breadth. Two hundred support tickets to triage; forty repositories to scan; ten unrelated leads to research. Independent items, no cross-item reasoning, results merged by concatenation rather than by judgment. This is why research systems fan out well — reading parallelizes — while tightly coupled artifact-building (a codebase, a document with one voice) resists decomposition.
3. The job genuinely doesn't fit. Sometimes the material truly exceeds one context budget. Then split — but along the seams where the least understanding has to cross the boundary. This is Parnas's information-hiding criterion for module boundaries, sixty years on, applying to prompts: decompose where the interface is thin, never where the org chart says a department should be.
4. Tiered capability. Mechanical breadth on a fast, cheap model; judgment calls on an expensive one. An economics criterion, and an honest one — as long as the boundary between “mechanical” and “judgment” is real and not aspirational.

“Because it feels more thorough” is not on the list. Rigor comes from the review step, not from headcount. A workflow doesn't become more serious by acquiring more boxes; it becomes more serious by being checkable.

Gates, not meetings

The reviewer split earns its keep only if the loop it creates is bounded. An unbounded refine-and-resubmit cycle is the agent equivalent of a recurring meeting: it consumes budget, oscillates between phrasings, and mistakes motion for convergence. A working gate returns a numbered fix-list against a concrete rubric, the producing step gets a fixed number of repair rounds — we use at most two — and then the loop stops regardless. If two evidence-backed repair rounds haven't converged, the defect isn't in the execution; it's upstream, in a misread intent, and no amount of additional rounds will fix that. Stop, re-read the brief, redesign.

Right-sizing as a discipline

The practical rule we run everything through: complexity must be derived from the structure of the job, never from the ambition of the builder. In our own system, the design step is required to state which of the four criteria forces each split, and the fallback is always the same — one excellent prompt. Watching real briefs flow through it, the distribution is humbling: most work right-sizes to a single prompt or a short pipeline with one gate. Orchestras are rare, and when they're real, the criterion is always breadth — never prestige.

None of this says multi-agent is wrong. Production systems that clear the bar exist, and when breadth is real the fan-out is glorious. The claim is narrower and, we think, more useful: splitting is a cost, defaulting to it is a smell, and the burden of proof always sits with the split. Start from one excellent prompt. Earn every box you add. And pay gladly for the one split that buys you fresh eyes — it's the only one that gets cheaper the more the others fail.