The Orchestra

Anthropic's practical guide treats orchestrator-workers as a workflow for tasks whose pieces cannot be listed in advance. Coding changes whose file list depends on the request, and search jobs whose sources depend on the question, are the examples they give. The topology looks like a team. The control path is still a lead that assigns, waits, and merges.

Anthropic's research system is the production version of that sketch: a lead plans, saves the plan, and spawns specialized subagents. The lead is also the place that can decide the work is done. That stop condition is part of the role. Early versions of that system spawned dozens of workers for simple queries; the fix was teaching the orchestrator how to delegate, not adding another layer of managers.

Cognition's critique still applies at this stand. If the lead hands out vague slices and never shares the decisions those slices depend on, the workers will invent incompatible worlds. The conductor is accountable for the brief. A one-line instruction such as "research the semiconductor shortage" is how you get three workers doing the same search.

The clean combination is breadth-first research. A question such as "list the board members of every Information Technology S&P 500 company" is a pile of independent lookups. Anthropic's lead could decompose that; a single sequential agent could not finish. That is a task-shape argument, not a maturity badge.

The same post is honest about the bill. Multi-agent systems in their data used about fifteen times the tokens of chat. They also say domains that require shared context, including most coding, are a poor fit today. If you quote 90.2% without 15×, you are not reading the score.

Delegation quality is the difference between an orchestra and a crowd. Anthropic had to specify objective, output format, tools, and boundaries after short instructions caused duplicated work. Human testers still caught a bias toward SEO content farms over primary sources. The research desk is only as good as the source rule in its brief.

Anthropic's agent guide still lists coding products that change many files as an orchestrator-worker example, because the file list is not known up front. That is compatible with a single builder under a lead. It is not a license for two builders to invent the architecture in parallel.

Cognition's June 2025 critique is the required counterweight. Sharing the original prompt is not enough once tool calls and multi-turn decisions have happened. Actions encode taste, naming, and structure. Two unseen builders will not converge on those. Their recommended default is a single-threaded agent; extra agents, if any, answer questions so the main trace stays short.

On this map, the builder is therefore a specialist with a tight contract: change the files named in scope, or justify an extra file in the patch notes. The verifier, not a sibling builder, is the combination that pays.

The verifier is how specialized agents stay honest. Anthropic's building-effective-agents post names evaluator-optimizer as its own workflow, useful when a human's feedback would help and a model can give that feedback. Translation and multi-round search are their examples. The two signs of fit are a clear rubric and measurable improvement.

The research system adds a later specialist: a citation agent that grounds claims after the lead synthesizes. That is a verifier with a narrow artifact. It exists because synthesis is where sources get smoothed away. If your multi-agent design has no equivalent pass, the lead is grading its own homework.

MetaGPT is on this stand as a warning, not as a product recommendation: cascading hallucinations are the failure mode of naive chaining. A verifier that cannot fail the previous artifact is another chain link.

OpenAI's builder guide is the other source on this stand: maximize one agent first, then split when prompts or tools become unmanageable, or when a manager or handoff pattern is required. Policy and tool overload are valid split reasons. Fashion is not.

The useful combination is boring. A classifier, a permissions check, a spend cap, or a human approval sits beside the worker. Anthropic notes that specialized attention on one consideration tends to beat a single overloaded call. That is sectioning, not a swarm.

On this map, policy is drawn in the danger color because a missing gate is how multi-agent systems become expensive and irreversible. The artifact is a decision record, not a vibe of carefulness.

Anthropic's research lead writes its plan to memory because a 200,000-token window will truncate, and the plan is not an optional souvenir. That is the difference between memory as a product feature and memory as a survival mechanism for long jobs.

Cognition's first principle belongs on the same stand: share context, and share full traces, not isolated messages. Their second principle explains why: actions carry implicit decisions. A memory that stores only the user's original sentence has already dropped the tool-call trail that interpreted it.

The Orchestra treats memory as the score. Research notes, patches, test reports, and policy decisions are staves. If you cannot point to the bar a later agent is playing from, you do not have a multi-agent system. You have several chats.

OpenAI's guide puts handoffs late: after a single agent is doing real work, and after prompt or tool complexity, manager patterns, or policy boundaries make a split rational. That order matters. A handoff graph drawn before a working single agent is choreography.

AutoGen is here because it is a research framework for composing conversable agents with models, tools, code, and humans. It is evidence that conversation can be a coordination medium. It is not evidence that conversation is sufficient. The useful residue is the same as everywhere else on this map: who can call tools, what they must return, and when a human enters.

Anthropic's research post notes that their lead still runs subagents synchronously, waiting on each set before continuing. Even a sophisticated production system has not made free-form agent chat the coordination layer. If you need asynchrony, you are taking on state consistency as a product problem, not unlocking a higher form of intelligence.