- Home
- AI agents
- Orchestration and multi-agent systems
- Multi-agent patterns
Build · Orchestration and multi-agent systems
Multi-agent orchestration patterns: which one fits your workflow?
Multi-agent patterns are the shapes a system takes when work is split across several AI agents: a supervisor that briefs workers and merges results, handoffs that pass a conversation along, a group chat with a manager, a swarm with shared memory, or a graph where developers fix the path. Each trades predictability and cost for breadth.
In this guide: 83 pages
Key takeaways
- A multi-agent system pays off on broad, parallel work with a high value per task. Anthropic's research system beat a single agent by 90.2% on its internal evaluation, at about 15 times the tokens of a chat.
- Supervisor and workers is the pattern to learn first. Handoffs suit conversations that change domain; group chat and swarm are rarely a first choice in an enterprise.
- Most of the production systems we see end up as a graph: developers fix the path and the model decides at a few marked nodes. Every major framework now ships one.
- The failures are predictable. Fragmented context, conflicting decisions and unverified results account for most of them, and all three are design choices.
What is a multi-agent system, and when does it pay off?
A multi-agent system splits one job across several agents, each with its own instructions, tools and context window, coordinated by a lead agent, a handoff rule or a graph. It pays off when the job divides into pieces that can run in parallel without seeing each other's work, and when the value per task covers the token bill.
The public measurement most often cited is Anthropic's Research feature (13 Jun 2025): a lead Claude Opus 4 agent with Claude Sonnet 4 subagents outperformed a single Opus 4 agent by 90.2% on Anthropic's internal research evaluation. The same note reports the price: agents use about four times the tokens of a chat and multi-agent systems about 15 times, and token usage alone explained 80% of the performance variance on BrowseComp. Breadth-first research, where each worker searches a different corner, is the case the pattern was built for.
The counter-case is as well documented. Cognition's note of 12 Jun 2025 argues that most coding and analysis tasks have fewer independent pieces than they appear to, that subagents which cannot see each other's traces make conflicting decisions, and that a single-threaded agent with context compression wins. Anthropic agrees that coding has fewer truly parallelizable tasks than research. OpenAI's guide to building agents says to get as far as possible with one agent before splitting.
Two tests we apply before splitting a task: can the lead write a brief a worker can act on without the lead's conversation, and can the lead verify the worker's output without redoing the work? If both hold, one of the patterns below will probably pay. If either fails, stay with one agent.
Which patterns exist, and how does each work?
Six patterns cover what the frameworks ship on 2 Nov 2026, plus the single agent they all start from. The names differ by vendor; the table uses the most common one and lists where each ships.
Multi-agent patterns and where they ship, as of 2 Nov 2026.
| Pattern | How it works | Where it ships |
|---|---|---|
| Supervisor (orchestrator-workers, manager, agents-as-tools) | A lead agent plans, writes a brief per worker, runs workers in parallel and merges the results; workers are invoked like tools and return to the lead | Anthropic Research; OpenAI Agents SDK and Strands agents-as-tools; Microsoft Agent Framework concurrent and Magentic-One, where a manager keeps a task ledger |
| Handoffs (decentralized) | Control and the conversation pass to another agent, which takes over; the first agent is out of the loop | OpenAI Agents SDK handoffs, which OpenAI describes as a specialized tool call for transferring control between agents; Microsoft Agent Framework handoff; Strands handoff_to_user, which hands to a person |
| Group chat | Several agents share one conversation; a manager chooses who speaks next, round by round | Microsoft Agent Framework (inherited from AutoGen, which is now archived); Foundry multi-agent workflow template (preview) |
| Swarm | Agents coordinate through shared memory and decide among themselves who acts; no fixed controller | Strands swarm |
| Graph (workflow) | Developers define nodes and edges; the model decides only at marked nodes; checkpoints and interrupts hang off the graph | LangGraph; Microsoft Agent Framework workflows; Strands graph and workflow (a fixed DAG); Google ADK 2.0 graph workflows; LlamaIndex Workflows |
| Remote agents over A2A | Agents run as separate services, publish an agent card at a well-known URL and delegate tasks over the A2A protocol (1.0.0, 12 Mar 2026) | Microsoft Foundry (A2A 1.0 generally available); Copilot Studio connected agents (A2A generally available Apr 2026); Google Agent Registry (A2A v1); Strands; ADK |
Google's ADK adds a vocabulary worth knowing: Sequential, Parallel and Loop workflow agents for the fixed parts and LLM-driven transfer for the judgment parts, with graph workflows and a human-input node in ADK 2.0 (Python 2.0 generally available 19 May 2026). Strands' rule of thumb covers the rest: graph when you want to define the flow, swarm when agents should coordinate on their own, workflow when the process is fixed.
How do you choose between them?
Choose by the shape of the work and by who must hold the context. Parallel, independent pieces want a supervisor; a conversation that changes domain wants a handoff; a process with a known shape wants a graph; shared context wants one agent.
- Supervisor when the lead can brief each worker in writing, the workers do not need each other's results, and the lead can verify what comes back. Research, document review at volume and content generation fit. Cap the fan-out and the tokens per worker.
- Handoffs when a user conversation moves between domains, such as triage to billing to returns, and the second agent needs the transcript rather than a brief. Pass the full context with the handoff, and log who handed off to whom.
- Graph when the process has a shape you can draw and a few steps need judgment. Most enterprise agents end up here, because approvals, durable waits and audit trails attach naturally to nodes.
- Group chat and swarm when you are exploring and the value of a surprising path outweighs predictability. Keep them away from anything that writes to a system of record until the path is understood and can be drawn as a graph.
- Remote agents over A2A when the agents belong to different teams or vendors and must stay separately deployed, versioned and governed. Inside one application, call sub-agents directly; A2A adds a network hop and a trust boundary you then have to manage.
Whatever the pattern, keep the following fixed: the lead or the graph holds the context of record, every worker runs under its own identity with narrower rights than the lead, and nothing a worker produces reaches a system of record without a verification step.
What goes wrong in multi-agent systems?
Most failures are design failures, and the public taxonomies agree on where they sit: in how the system was specified, in how agents misalign with each other, and in how little the results are verified.
The MAST study (17 Mar 2025, "Why Do Multi-Agent LLM Systems Fail?") catalogued 14 failure modes across popular frameworks and grouped them into three categories: system design and specification, inter-agent misalignment, and task verification. Its authors note that performance gains on popular benchmarks are often minimal. In our reviews the three categories map onto three missing artefacts: a written brief per worker, a shared trace, and a verification step.
Cognition's version is concrete: two subagents asked to build a Flappy Bird clone produced a background in a Super Mario style and a mismatched bird, because actions carry implicit decisions and the agents never saw each other's decisions. Its two principles, share context with full agent traces rather than individual messages, and avoid parallel agents that make conflicting decisions, make a usable short checklist.
Then the operational failures. Errors compound, so a worker that misreads the brief at step two poisons every step after it; Anthropic's fix was checkpoints with resume-from-failure and rainbow deployments so a release leaves in-flight agents alone. Cost runs away when fan-out is uncapped. And every worker with tools is another surface for prompt injection and over-broad permissions, which is why a worker needs its own scoped identity and the same guardrails as a user-facing agent. The guardrails and identity topics cover both.
How do Microsoft, AWS and Google implement it?
Where each pattern lives on the three clouds, with statuses as seen on 2 Nov 2026.
Capability, Microsoft, AWS, Google
| Capability | Microsoft | AWS | |
|---|---|---|---|
| Supervisor and workers | Agent Framework concurrent and Magentic-One; Copilot Studio connected agents | Strands agents-as-tools on AgentCore Runtime; the AgentCore harness alone covers multi-agent only partly (AWS) | ADK sub-agents and Parallel agents on Agent Runtime |
| Handoffs | Agent Framework handoff; Copilot Studio transfer to a person or a connected agent | Strands handoff_to_user; OpenAI Agents SDK handoffs hosted on AgentCore Runtime | ADK LLM-driven transfer between sub-agents |
| Graph | Agent Framework workflows with checkpoints; Foundry multi-agent workflows (public preview since 25 Nov 2025) | Strands graph and workflow; Step Functions around the agents | ADK 2.0 graph workflows with a human-input node (GA 19 May 2026) |
| Remote agents over A2A | Foundry A2A 1.0 GA; Copilot Studio A2A GA Apr 2026 | Strands and AgentCore A2A; the documented path pins the 0.3 SDK | Agent Registry with A2A v1 agent cards (GA 18 Jun 2026) |
Foundry's visual multi-agent workflows were public preview when announced; we could not confirm a later status on 2 Nov 2026.
Frequently asked questions
What is the difference between a supervisor and a handoff?
In a supervisor pattern the lead stays in control: it briefs a worker, waits, and merges the result, and the worker never talks to the user. In a handoff the second agent takes over the conversation and the first is out of the loop. OpenAI's guide calls these the manager and decentralized patterns.
Is a multi-agent system more accurate than one agent?
Sometimes, on broad parallel work. Anthropic's research system outperformed a single agent by 90.2% on its internal evaluation, but the MAST study found gains on popular benchmarks are often minimal, and Cognition found single agents win on coupled tasks. Measure on your own evaluation set before deciding.
Which multi-agent framework should we use?
The one your cloud hosts natively, with the loop in your own code: Microsoft Agent Framework on Azure (1.0 generally available 3 Apr 2026), Strands or LangGraph on AWS, Google ADK on Agent Runtime. All three ship supervisor, handoff and graph patterns. AutoGen is archived, and Semantic Kernel became Agent Framework.
Do workers need their own identity?
Yes. A worker that inherits the supervisor's rights is an over-permissioned agent with its own prompt-injection surface. Give each worker an identity with narrower rights, log its tool calls under that identity, and put the same guardrails on it as on a user-facing agent. The identity topic shows how on each cloud.
More in Orchestration and multi-agent systems
Sources
Each source was read on 2 November 2026. The date shown is the date on the page, or the page’s ms.date for Microsoft Learn; an undated page shows the date it was read.
- How we built our multi-agent research system
- Don't Build Multi-Agents
- Why Do Multi-Agent LLM Systems Fail? (MAST)
- A practical guide to building agents (PDF)
- openai-agents on PyPI
- Introducing Strands Agents 1.0
- Strands multi-agent patterns
- Introducing Microsoft Agent Framework
- Microsoft Agent Framework version 1.0
- Introducing multi-agent workflows in Foundry Agent Service
- What's new in Copilot Studio
- Agent Development Kit: easy to build multi-agent applications
- adk-python v2.0.0 release
- Amazon Bedrock Agents Classic maintenance mode
- A2A protocol specification, version 1.0.0
Planning an AI program?
Talk to the team that builds it and runs it after launch.
- Built to SOC 2, HIPAA and GDPR standards
- EU AI Act-aligned delivery
- Microsoft Solutions Partner