← Library · Concept
Multi-agent systems — when and why to split work between roles
A multi-agent system divides work among specialized agents. It can help when one agent struggles with complex rules or tools, but it adds coordination, cost, and new places for errors.
First, check whether one agent is enough
A multi-agent system is a workflow in which several agents pass parts of the work to one another. It is not an automatic improvement. OpenAI recommends expanding a single agent’s capabilities first, then adding agents when instructions become too branched or the agent repeatedly confuses similar tools.
A simple request to summarize a document does not need three roles. A process requiring separate research, source checking, and an approved action may be clearer when responsibilities are split.
Two common patterns
In the manager pattern, a primary agent delegates sub-tasks to specialists and combines their output. It fits workflows where the user should still interact with one point and the primary agent retains the decision and context.
In the handoff pattern, an agent transfers a task when it reaches another agent’s domain. It fits clearly separated conversational roles. OpenAI describes these as manager and decentralized handoffs. The Agents SDK documentation shows that orchestration must explicitly specify who controls the run and who receives the result.
Split work, not accountability
Give each role an input, allowed tools, expected output, and exit condition. One agent may gather material, another compare it with sources, and a third prepare a proposal. One workflow owner must still decide whether the result is sufficient and whether an external action may occur.
For example, a customer-support agent sends an invoice to an accounting specialist. The specialist returns a structured finding, not a customer message or a refund command. The primary agent checks the policy and routes a sensitive refund decision to a person.
Cost and security are central
Each handoff adds context, a model call, and a chance to lose information. Compare the entire task: quality, duration, cost, and failures—not just each role’s capability. Splitting work can improve clarity and testing, but for a small task it can cost more and be less reliable than one well-configured agent.
Shared tools and data also need clear ownership and narrow permissions. OWASP notes that excessive functionality, permissions, or autonomy can enable damaging actions after unexpected or manipulated output. Its Excessive Agency guidance supports reducing capability scope and protecting sensitive actions.
When not to use it
Do not add roles merely because they sound advanced. If you cannot state each role’s benefit, input, and measurable output, start with one agent. A multi-agent system earns its complexity when it simplifies decisions or safely separates work—and you can verify that on real tasks.