Multi-agent systems, the fashionable architecture of the moment, come in three varieties, and none of them answers the question a patient person might ask first. They are hand-designed, born one-shot from a generator, or evolved by code search with no constraints. In each case the number of agents is settled before the work begins. The question the designers never ask is how many agents the task wants, because the design already knows.
A team of Amazon scientists — Amit Dhanda, Sina Amini Niaki, Anish Acharya, Brian Verkhovsky — has now asked it. In a paper posted to Amazon’s science site in September 2026, they describe OrchOpt, a system that treats agent count itself as a searchable parameter, subject to what they argue is the same discipline applied to neural architectures: bounded edits, held-out validation gating, and a search that can move in both directions. Add an agent when an agent is needed. Remove one, when one is not.
The removal is the interesting part, and the system is built to make it cheap. A proposer model reads failure analysis and proposes small, bounded edits — merge, split, skill — to the current architecture. An edit is accepted only when it strictly improves performance on held-out validation. Nothing is adopted on faith. The mechanism that makes subtraction possible is what the authors call knowledge preserving merge: when two agents are folded into one, the absorbed agent’s expertise does not disappear. It becomes an activatable skill. The search is fully reversible, they write, at zero additional inference cost — meaning a wrong turn can be un-turned without paying to run two agents where one would do.
What this means in practice is that a single agent and a multi-agent committee can be made to compete directly for the same job, in the same arena, under the same rules. The committee does not win by default.
The proof was run across nine benchmarks: math, knowledge question-answering, code generation, reading comprehension, multi-hop reasoning, negotiation, research. The optimizer was initialized in every case with the same baseline design and then left to decide for itself. It decided, on five of the nine domains, to simplify down to one agent. On the other four it elaborated outward, to two or three.
Five times it subtracted. Four times it added. That is to say, the freedom to remove was exercised more often than the freedom to build.
The gains, as the authors report them, were not marginal. The found architectures scored 20 percentage points better than the ADAS system on MGSM, and 16.7 points better than AFlow on HotpotQA. On Bargaining, the negotiation benchmark, they exceeded MARBLE’s published-best topology by 27.5 points. The searches converged quickly — in between 7 and 42 candidate evaluations, roughly one to three dozen drawings before the picture was final.
It is worth holding the numbers the way the paper holds them: as the outcome of a competition in which the extravagant option lost often. The field’s default has been addition — an agent for each subtask, a committee for each problem, on something like the assumption that more minds make a better mind. What OrchOpt found, on its own recorded validation, is that the assumption fails more than half the time. Sometimes the committee is one person.
The authors’ claim is modest in its architecture and large in its implication: that agent count and coordination topology are not design intuitions but parameters, and parameters can be searched. The search they built can discover what they say no existing method can — that a single agent is optimal.
A single agent is optimal. It reads, in the paper’s flat phrasing, less like a finding than a correction.