--- schema: formation.pattern/v0.1 kind: pattern visibility: public aliases: - 'orchestrator-worker (Anthropic)' - orchestrator-workers - 'supervisor agent (LangGraph)' - 'manager agent' - 'hierarchical process (CrewAI)' - 'lead agent with subagents' canonical_url: https://topologyindex.com/patterns/supervisor.md description: 'One controlling role directs subordinate workers and decides what happens next.' executable_here: false executor_requirements: - agent_to_agent_messages - routing_on_unmeasured_conditions family: hierarchy findings_page: /patterns/supervisor/findings.md observed_in: - /research/openai-hugging-face-agent-coordination.md path: /patterns/supervisor.md pattern: supervisor pattern_index: /patterns/index.md product_api_version: v1 published_findings: - benchmarks: [] compared_against: 'A single agent using the same model as the lead agent' direction: helped source_id: web:anthropic.com/engineering/multi-agent-research-system task_domain: 'Open-ended web research' task_domains: - research url: https://www.anthropic.com/engineering/multi-agent-research-system - benchmarks: - GAIA - AssistantBench - WebArena compared_against: 'Published state-of-the-art agent systems on each benchmark' direction: no_clear_gain source_id: arxiv:2411.04468 task_domain: 'Generalist web, file and coding tasks' task_domains: - coding - operations url: https://arxiv.org/abs/2411.04468 - benchmarks: - Finance-Agent - PlanCraft - BrowseComp-Plus - Workbench - 'SWE-bench Verified' - Terminal-Bench compared_against: 'Single-agent systems and independent, decentralized and hybrid multi-agent architectures' direction: mixed source_id: arxiv:2512.08296 task_domain: 'Agentic benchmarks spanning finance, web browsing, planning, workplace, software and terminal tasks' task_domains: - coding - research - reasoning - operations url: https://arxiv.org/abs/2512.08296 - benchmarks: [] compared_against: 'Single-agent approaches' direction: helped source_id: arxiv:2412.05449 task_domain: 'Enterprise assistant scenarios' task_domains: - operations url: https://arxiv.org/abs/2412.05449 - benchmarks: - MAST-Data compared_against: 'Expected task success of the same frameworks' direction: hurt source_id: arxiv:2503.13657 task_domain: 'Coding, math and general agent tasks across several open-source multi-agent frameworks' task_domains: - coding - reasoning url: https://arxiv.org/abs/2503.13657 - benchmarks: - AIxCC compared_against: 'A single-agent patcher, a fixed workflow and a general-purpose coding agent on the same tasks' direction: mixed source_id: arxiv:2603.01257 task_domain: 'Security: repairing real-world vulnerabilities in large Java projects' task_domains: - coding - security url: https://arxiv.org/abs/2603.01257 qualifying_evidence: [] references: - https://www.anthropic.com/engineering/building-effective-agents - https://www.anthropic.com/engineering/multi-agent-research-system - https://arxiv.org/abs/2308.08155 schema_version: v0.1 title: 'When to use the supervisor multi-agent pattern (orchestrator-worker, orchestrator-workers)?' unmet_executor_requirements: - agent_to_agent_messages - routing_on_unmeasured_conditions --- # When to use the supervisor multi-agent pattern (orchestrator-worker, orchestrator-workers)? **Short answer:** start with it when the task has subtasks unknown until it starts; avoid it when a fixed plan would do. A hypothesis from the [decision guide](/patterns/index.md), not a ranking; [published findings](/patterns/supervisor/findings.md) keep the unfavourable ones. The `supervisor` multi-agent pattern (family `hierarchy`): one controlling role directs subordinate workers and decides what happens next. Also called: orchestrator-worker (Anthropic), orchestrator-workers, supervisor agent (LangGraph), manager agent, hierarchical process (CrewAI), lead agent with subagents. ## The arrangement - Structure: A supervisor role plus workers that act only when directed. - Control: The supervisor assigns work, inspects what comes back, and redirects or stops. - Shared state: Whatever the supervisor passes down; workers do not address each other. ## When it is hypothesised to fit Hypothesised to fit when the work needs one place that holds the plan, and the workers can be narrower and cheaper than the role directing them. A hypothesis to test on a workload, not a finding: this service has not tested it. ## Failure modes to watch for - Every result is filtered through one context window, which is both a bottleneck and a single point of failure. - A supervisor that is itself a model can be argued out of its own instructions by the work it is reading. - Subordinates report upward as summaries, so the supervisor sees a tidier run than the one that happened. These are things to watch for, not outcomes anyone measured here. ## What published studies found Attributed to each source and stated without figures; unfavourable results are included on purpose, and none of this is evidence produced by this service. What each was compared with is in `published_findings`; each finding in words, with its caveat, is at [/patterns/supervisor/findings.md](/patterns/supervisor/findings.md). Reviewed 2026-09-23. Each label says how this pattern fared against what it was compared with, as the source reports it; the labels are defined at [/docs/schemas/pattern/v0.1.md](/docs/schemas/pattern/v0.1.md). - **Helped**: [Anthropic Engineering: How we built our multi-agent research system](https://www.anthropic.com/engineering/multi-agent-research-system) - **No clear gain**: [Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks](https://arxiv.org/abs/2411.04468) - **Mixed**: [Towards a Science of Scaling Agent Systems](https://arxiv.org/abs/2512.08296) - **Helped**: [Towards Effective GenAI Multi-Agent Collaboration: Design and Evaluation for Enterprise Applications](https://arxiv.org/abs/2412.05449) - **Hurt**: [Why Do Multi-Agent LLM Systems Fail?](https://arxiv.org/abs/2503.13657) - **Mixed**: [A Systematic Study of LLM-Based Architectures for Automated Patching](https://arxiv.org/abs/2603.01257) ## Sources that describe it Where the arrangement is described. A source is a reason to test it, never a result for it. - [Building effective agents (Anthropic)](https://www.anthropic.com/engineering/building-effective-agents) - [How we built our multi-agent research system (Anthropic)](https://www.anthropic.com/engineering/multi-agent-research-system) - [AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation (Wu et al.)](https://arxiv.org/abs/2308.08155) ## Can this deployment execute it No. The runner does not provide everything this arrangement needs. What is missing: - `agent_to_agent_messages` (a channel on which one agent addresses another): The adapter sees a single role invocation, and the tool surface offers no channel to another agent. - `routing_on_unmeasured_conditions` (the next role chosen from conditions the executor does not measure): A condition outside the measured signal set is refused as an unsupported capability, and every adaptation rule in an execution lock is refused with it. Supervision itself is not missing here: the runner supervises every episode, asks a guard before each invocation and holds the bounds. What cannot run is supervision by a role, where a model decides what another agent does next. A topology formation that declares what this deployment cannot run is still stored, and is refused with an unsupported-capability error rather than reduced to a simpler arrangement and run anyway. ## Evidence for this pattern None. No qualifying evidence exists here for this pattern, and by design none ever will: an evidence snapshot is about the complete execution configuration that produced the observations, so the most it can support is a claim about that configuration on that workload. ## Observed in Reports that describe this arrangement. A report is a reason to test an arrangement, never a result for it. - [Agent coordination in the OpenAI and Hugging Face evaluation incident](/research/openai-hugging-face-agent-coordination.md) (descriptive_only) ## How evidence works here A pattern names how work is organized, never how well that works. Evidence in this service attaches to a complete execution configuration — a topology formation together with an execution lock pinning models, tools and runtime — and never to a pattern on its own, so no page here carries a success rate, a cost or a run count for an arrangement.