Multi-Agent Collaboration
What It Is
Section titled “What It Is”Multi-agent collaboration is a pattern where multiple AI agents — each with different roles, tools, or expertise — work together to accomplish a task. Instead of one general-purpose agent doing everything, the work is divided among specialists that communicate, delegate, and coordinate.
Think of it like a team at a company: a researcher gathers information, an analyst processes it, a writer drafts the report, and an editor reviews it. Each agent focuses on what it does best, and the combined output is better than any single agent could produce alone.
Why It Matters
Section titled “Why It Matters”Single agents hit a complexity ceiling. As tasks grow more complex, a single agent’s context window fills up, its instructions become contradictory, and its performance degrades. Multi-agent systems solve this by decomposing the problem across agents that each operate with focused context and clear responsibilities.
Multi-agent collaboration also enables separation of concerns — a safety-critical agent can enforce policies while a creative agent generates content, without either interfering with the other’s role.
How It Works
Section titled “How It Works”┌──────────────┐│ Orchestrator ││ (coordinator)│└──┬───┬───┬───┘ │ │ │ ▼ ▼ ▼┌────┐┌────┐┌────┐│ A1 ││ A2 ││ A3 ││ ││ ││ │└────┘└────┘└────┘Researcher Analyst WriterCommon ways to arrange the team, each with its own page in the Workflow Architecture Patterns catalogue:
- Orchestrator-Workers — a central agent assigns tasks to specialist agents, collects results, and synthesizes the final output.
- Prompt Chaining — agents in a sequence, where each agent’s output becomes the next agent’s input (researcher → analyst → writer).
- Group Chat — agents argue different perspectives in a shared conversation, and a moderator synthesizes the result.
- Swarm — agents work as peers with no lead, each contributing to a shared workspace.
The catalogue also covers Handoff, Parallelization and the rest, maps each pattern to an autonomy level, and has a decision tree for choosing one.
The key design decisions are:
- How many agents? — Use the minimum number needed. More agents means more coordination overhead: in Anthropic’s own data, a multi-agent system used about 15 times the tokens of a chat, against about 4 times for a single agent (How we built our multi-agent research system, June 13, 2025).
- How do they communicate? — Through shared context, message passing, or a shared workspace.
- Who decides what? — An orchestrator, a round-robin, or autonomous negotiation.
Example
Section titled “Example”Customer exchange scenario
Section titled “Customer exchange scenario”A customer exchange can be handled by a team of agents:
- Triage Agent — Reads the customer’s request, classifies it as “exchange,” and routes it to the right specialist.
- Order Agent — Looks up the order, verifies return eligibility, and handles the logistics.
- Inventory Agent — Checks stock levels and finds the best warehouse to ship from.
- Communications Agent — Drafts the customer-facing email with exchange details and return instructions.
Each agent has access only to the tools it needs — the Communications Agent can’t process payments, and the Order Agent can’t send emails. This limits the blast radius of any single agent’s mistakes.
Content production
Section titled “Content production”A marketing team’s content pipeline as a multi-agent system:
- Research Agent — Gathers data, competitor analysis, and market trends
- Writer Agent — Produces a draft based on the research brief
- Editor Agent — Reviews for clarity, accuracy, and brand voice
- SEO Agent — Optimizes headlines, metadata, and keyword placement
The Writer Agent never sees raw data — it receives a structured brief from the Research Agent. The Editor Agent doesn’t know about SEO — it focuses purely on quality. The separation keeps each agent’s instructions and context on one concern, instead of one agent juggling all four.
When to Use It
Section titled “When to Use It”- Tasks that require multiple distinct skill sets (research + writing + analysis)
- Workflows where separation of concerns improves quality or safety
- Problems where debate or multiple perspectives lead to better answers
- Systems that need to scale — adding a new capability means adding a new agent, not rewriting the existing one
- Production systems where different agents need different permission levels
Related Patterns
Section titled “Related Patterns”- Planning — An orchestrator agent often uses planning to coordinate the team
- Tool Use — Each agent typically has its own set of tools
- Reflection — A critic agent reviewing another agent’s work is multi-agent reflection
- Guardrails — Each agent can have its own guardrails, limiting what it can do
- Agent Capability Patterns
Further Reading
Section titled “Further Reading”- Andrew Ng — Agentic Design Patterns Part 5, Multi-Agent Collaboration (The Batch, April 17, 2024) — the last of his four patterns
- Wu et al. — AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation (arXiv, August 16, 2023)
- Anthropic — How we built our multi-agent research system (June 13, 2025)
- Anthropic — Building effective agents (December 19, 2024)
- CrewAI — Processes (v1.15.30 docs, accessed 2026-10-24) — sequential and hierarchical crews