The reason coordination becomes the expensive part sits in Anthropic's description of the work itself. It says research is open-ended and path-dependent, that no fixed path can be hardcoded, and that a linear one-shot pipeline cannot handle these tasks [5]. An organisation that cannot specify the steps also cannot test the steps. It can only evaluate outcomes, and it has to make a system that wanders recover instead of fail, which is a budget line for evaluation harnesses and for the engineers who read transcripts rather than for model procurement [4].
The token arithmetic makes the trade concrete. Anthropic reports one multiplier for agents against chat and another for multi-agent systems against chat; fifteen over four puts a multi-agent run at roughly 3.75 times the inference of a single agent [14][12]. So the configuration that produced the reported eval advantage costs close to four single-agent runs per query [7]. Anthropic says these architectures burn through tokens fast and that economic viability depends on the task, in a sentence the available text cuts off mid-clause [13].
A skeptic reads the same post and says model capability is still the binding constraint, citing Anthropic's own line that upgrading to Claude Sonnet 4 is a larger gain than doubling the token budget on Sonnet 3.7 [11]. Both readings hold. On BrowseComp, which tests whether browsing agents can locate hard-to-find information, token usage alone explains 80 per cent of variance, leaving about fifteen points of the explained variance for tool calls and model choice together [16][10][15]. The model sets the yield per token; the harness decides how many tokens a run gets and whether the run finishes. Model upgrades also arrive on the vendor's calendar, while the harness is what a team can staff this quarter.
What the record does not say is worth stating plainly. This is Anthropic's own engineering account of its own product [1], and the headline comparison rests on an internal eval whose contents are not published [7]. We do not know its sample size or task mix, and the post claims the multi-agent advantage especially for breadth-first queries that pursue several independent directions at once [8]. The board-deck version of this is "multi-agent beats single-agent by 90 per cent"; the incomplete part is that a single illustration about identifying board members across the Information Technology companies in the S&P 500 [9] is a demonstration, not a benchmark, and depth-first behaviour is not characterised at all.
Adopt the architecture and you inherit its shape. Parallel subagents with their own context windows, tools and prompts buy separation of concerns and less path dependency [6], and they also mean concurrency ceilings and a failure surface that only becomes visible across many runs, which is the reliability engineering the post names as a new challenge [4]. That is this quarter's decision. Next quarter's consequence is procurement and sign-off: the question stops being which model to call and becomes how much search per query the business will fund, and who owns that number when one research task consumes what fifteen chat sessions would [12].
