The brief and the check

Once several AI systems can each do real work, it is natural to picture the person in charge as a dispatcher: someone who looks at a stream of tasks and sends each to the right worker, the fast one, the careful one, the one that is good at code. On that picture the scarce skill stops being doing the work and becomes routing it, and "managing a fleet of agents" becomes a job.

The picture is half right. The work does move away from doing, but it does not settle at routing. Choosing which model gets which task is the part that is already being automated. What remains hard, and what fleets fail at most visibly, are the two ends: writing the brief that tells each agent what its piece is, and checking what comes back with something that does not share its mistakes.

The workers are less specialized than the image suggests. The best-documented working fleet is Anthropic's research system, in which a lead agent running Claude Opus 4 plans an investigation and launches subagents running Claude Sonnet 4 to search in parallel. On the company's own internal research evaluation it beat a single Opus 4 agent by 90.2%. Anthropic's own explanation is not specialization. It says such systems "work mainly because they help spend enough tokens to solve the problem", and on a public browsing benchmark token usage alone explained 80% of the variation in performance.¹ The subagents are instances of one model, each with its own context window, which the company describes as adding capacity for parallel reasoning.¹

That gain has a price and a shape. Multi-agent runs used about 15 times as many tokens as an ordinary chat, and the company says they suit valuable, breadth-first tasks, not work where agents must share context or depend on each other; most coding, it notes, has fewer truly parallel pieces than research.¹ A fleet of this kind is closer to hiring more of the same person for an afternoon than to assembling a team of specialists.

Where models really do differ, mostly in price and strength, the choice between them is turning into software. RouteLLM, a framework from UC Berkeley researchers and collaborators, trains a small router to decide, query by query, whether a cheap model is good enough or the expensive one is needed. On public benchmarks its routers cut cost by more than half without losing response quality, and kept working when routing between models they had not been trained on.² Inside Anthropic's system, the dispatching of subagents is done by the lead agent.¹ The dispatcher's seat is being filled by a model.

The brief is where fleets break. Anthropic reports that when its lead agent gave short instructions such as "research the semiconductor shortage", subagents misread the task or ran the same searches as one another; one studied the 2021 car-chip crisis while two others duplicated work on 2025 supply chains. Early versions spawned 50 subagents for simple questions. The fix was a fuller brief for each subagent: an objective, an output format, guidance on tools and sources, and clear boundaries.¹ Cognition, which builds a coding agent, describes the same failure in software. Split "build a Flappy Bird clone" into a background and a bird, and two subagents working from partial context can produce pieces in incompatible styles, because every action carries decisions the other agent never saw. Its default advice is a single agent with one continuous context.³

A study of more than 1,600 traces from seven multi-agent frameworks found that performance gains on popular benchmarks are often minimal, and sorted the failures into three groups: flaws in the system's design and task specification, misalignment between agents, and failures of verification.⁴ Misalignment between agents is the hand-off problem, and verification is the check.

The check is harder than it looks, because the obvious checker is another model. When two language models both answer a question wrongly, they often give the same wrong answer: on one large leaderboard, 60% of the time. Larger and more accurate models have highly correlated errors even across different architectures and companies. When one model is used to grade another, it inflates the accuracy of models weaker than itself, because it accepts the wrong answers it would have given.⁵ A fleet that checks its own work with a sibling can turn a shared blind spot into a confirmed result. The same study of failures found that explicit verifier agents help but are not a cure.⁴

A better model for the human job is the expediter in a restaurant kitchen, whose work that matters here happens at the two ends: the ticket, called clearly enough that the plates match, and the inspection of each plate at the pass before it goes out. Routing is becoming a feature. Writing a brief that divides work without leaving gaps or contradictions, and judging the result with knowledge that the workers do not share, are not, and they are the parts of the job that the fleet itself cannot supply.

Whether that stays true depends on a question the evidence leaves open: if fleets were built from genuinely different models, trained on different data by different people, would their errors become independent enough to check one another? The trend so far, toward the strongest models making the same mistakes, points the other way.