Supervisor–Worker Orchestration

In a supervisor–worker pattern, one component decomposes work, assigns subgoals, tracks status, and integrates outputs. Workers operate on bounded tasks with explicit inputs, tools, budgets, and deliverables.

The supervisor should coordinate state and acceptance tests, not simply judge polished worker messages. Its own model output remains fallible. It needs verifiable artifacts.

One system made exactly that substitution and reported what it bought. MetaGPT, published in 2024, took dialogue out of the channel between agents: “Unlike ChatDev (Zhao et al., 2023), agents in MetaGPT communicate through documents and diagrams (structured outputs) rather than dialogue. These documents contain all necessary information, preventing irrelevant or missing content.”

The numbers for that swap sit in the paper's SoftwareDev table. Executability 3.75 against ChatDev's 2.25. Human revision cost 0.83 against 2.5. Productivity 124.3 tokens per line of code against 248.9. On code benchmarks the same paper reports 85.9% Pass@1 on HumanEval and 87.7% on MBPP. Artifacts beating conversation is not a preference here. It is a comparison someone ran and scored.

Work leaves the Task allocator, passes through a Worker, is recorded in the Progress ledger, and is checked by the Artifact verifier. The Artifact verifier and the Integrator should never be the same team, and should never share a test.

The ledger box in that diagram is not a metaphor. Microsoft Research built one. Magentic-One, published in November 2024, implements the supervisor as two nested loops over two named ledgers: “At a high level, the workflow contains two loops, the outer loop maintains the task ledger, which contains the overall plan, while the inner loop maintains the progress ledger, which directs and evaluates the individual steps that contain instructions to the specialized agents.”

The outer loop's Task Ledger holds given facts, facts to look up, facts to derive and educated guesses, plus the plan. The inner loop's Progress Ledger is rebuilt by the Orchestrator every turn. It rebuilds it by answering five fixed questions. “Is the request fully satisfied (i.e., task complete)?” “Is the team looping or repeating itself?” “Is forward progress being made?” “Which agent should speak next?” “What instruction or question should be asked of this team member?” A stall counter — threshold ≤ 2 in their experiments — breaks the inner loop and forces re-planning.

Notice what the five questions never ask. None of them asks a worker whether it is confident. They are questions the coordinator answers about the record, on a fixed schedule. Only the answer to the first one is entitled to end the task.

A Central supervisor, Hierarchical supervision, and a Deterministic scheduler differ in how much of the coordination is a model decision. The more of that decision sits with the model, the more the design depends on what a worker is obliged to hand back. Magentic-One's Orchestrator is the central case in its plainest form: one coordinator holds both ledgers and decides, each turn, which agent should speak next. Success looks like a worker returning its result plus three things — the evidence behind it, the scope it actually covered, and the questions it could not settle. The failure to watch for is quieter. A supervisor approves incomplete work because its own context hides the branches that never reported.

A review of this pattern has four things to look at, and they fail in different ways. Start with what comes back: workers return artifacts with evidence, scope, and unresolved issues, or they return a claim of success with nothing attached to it. Check separately that the supervisor can distinguish blocked, partial, and complete work, since the three call for different next moves. Then put the last two under repeated or adversarial cases.

Signal 3 has a written standard behind it, so it can be tested against a numbered clause instead of a feeling. RFC 9110, the IETF's HTTP Semantics specification of June 2022, defines the property in terms of effect rather than response: “A request method is considered "idempotent" if the intended effect on the server of multiple identical requests with that method is the same as the effect for a single such request.” The same clause, §9.2.2, adds a rule about retrying. A client SHOULD NOT automatically retry a request with a non-idempotent method “unless it has some means to know that the request semantics are actually idempotent, regardless of the method, or some means to detect that the original request was never applied”. That is the precondition a supervisor needs before re-dispatching a worker task that touched the world. One of the two, in writing, before the retry.

Signal 4 needs repetition, and there is a published metric for it. τ-bench, published in June 2024, grades an agent by comparing the final database state against an annotated goal state. Its pass^k metric requires the same task to succeed on all k i.i.d. trials. The abstract gives the result: “Our experiments show that even state-of-the-art function calling agents (like gpt-4o) succeed on <50% of the tasks, and are quite inconsistent (pass^8 <25% in retail).” Behind the abstract: 48.2% average at pass^1 — 61.2% on τ-retail, 35.2% on τ-airline. In τ-retail the score falls from about 61% at pass^1 to below 25% at pass^8, a roughly 60% relative drop that only repeated runs expose. A supervisor that accepted the first trace saw the 61% and never met the other number.

One supervisor assignment, written out in full against a real workflow, is more revealing than a diagram of ten. Write down what the supervisor sends the worker, what it expects back, and what it will do with a partial answer. Keep writing until the weak point shows: the moment where the supervisor would sign off on work whose missing branches never reached its context.

Two of the five steps do not have to be invented. The status vocabulary in “Specify status updates” — pending, running, blocked, partial, failed, complete — already exists in ratified form. Google announced the Agent2Agent (A2A) protocol in April 2025 and donated it to the Linux Foundation that June. The specification stabilised at v1.0 in March 2026. It defines the unit of delegation directly: “Task is the core unit of action for A2A. It has a current status and when results are created for the task they are stored in the artifact.” The lifecycle enumerates nine TaskState values: TASK_STATE_UNSPECIFIED, SUBMITTED, WORKING, INPUT_REQUIRED, AUTH_REQUIRED, COMPLETED, FAILED, CANCELED and REJECTED. Four are marked terminal: COMPLETED, FAILED, CANCELED and REJECTED. Two are marked interrupted: INPUT_REQUIRED and AUTH_REQUIRED. Results are not prose but typed Artifact objects attached to the task. That is Signal 1 written as a wire format.

“Revalidate the root” has a measured before and after. In the MAST intervention study, ChatDev's topology changed from a DAG to a cyclic graph that “terminates only when the CTO agent confirms that all reviews are properly satisfied”. ProgramDev-v0 accuracy rose from 25.0% to 40.6%, +15.6 points, with HumanEval moving 89.6% → 91.5%. The paper states it plainly: “We demonstrate this in an intervention study where adding a high-level task objective verification step to ChatDev yields a +15.6% improvement in task success on ProgramDev (details in Appendix H).” The change was not a better model or a better prompt. It moved the decision to stop to a named check against the original objective.

Keep coordination state outside supervisor prose. A deterministic ledger should remain authoritative when workers or models change. The reason is the failure this pattern is prone to: a supervisor can approve incomplete work because its context hides missing branches, and a summary written by that supervisor will not show the gap.

The ledgers have a price tag. Ablated on the GAIA validation set, they cost real accuracy: “We find that the Orchestrator’s ledgers are important to Magentic-One’s performance: without the full ledgers, performance drops by 31%.” Removing any single worker agent cost between 21% (Coder, Executor) and 39% (FileSurfer). So the bookkeeping the coordinator does is worth about as much as a whole specialist worker. That is not what a reader expects of a component that produces no output of its own.

Keep the ceiling in view while designing against it. Magentic-One with GPT-4o and o1-preview reached 38.00 ± 5.5 on the GAIA test set and 27.7 ± 6.5 accuracy on the AssistantBench test set. With GPT-4o alone it reached 32.8 ± 3.2 on WebArena, a benchmark with no common test split, where the authors report all 812 tasks. Human performance on GAIA is 92.00 ± 3.1. So a reviewer goes to the ledger before accepting the summary. Ask what each worker actually handed back — with what evidence, over what scope, and with what left unresolved.