The decentralized pattern is a multi-agent orchestration pattern in which agents operate on equal footing and hand off workflow execution to one another based on their specializations. A handoff is a one-way transfer: the agent that receives it takes over execution and interacts with the user directly, rather than returning a result to whoever called it. It is one of two multi-agent categories A practical guide to building agents names as broadly applicable, the other being the Manager Pattern.
Usage
Modelled as a graph with agents as nodes, the decentralized pattern’s edges are handoffs that transfer execution, where the manager pattern’s edges are tool calls that return. In the OpenAI Agents SDK a handoff is itself a type of tool or function: when an agent calls it, execution begins immediately on the agent handed off to, and the latest conversation state is transferred with it (see OpenAI Agents SDK).
The guide’s worked example is a customer service workflow covering both sales and support. A
triage agent acts as the first point of contact, assessing queries and directing them to the right
specialist, with technical support, sales assistant and order management agents listed in its
handoffs array; a question about a recent purchase causes the triage agent to hand off to the
order management agent, transferring control to it. The guide notes the receiving agent can
optionally be equipped with a handoff back to the original agent.
When It Applies
The guide presents the pattern as optimal where no single agent needs to maintain central control or synthesize results — where it is acceptable, and preferable, for each agent to take over execution and deal with the user itself. It names conversation triage as the archetypal case, and more generally any scenario where specialized agents should fully take over certain tasks without the original agent needing to stay involved.
It assumes a runtime that can carry conversation state across the transfer, since a handoff moves the interaction rather than copying a task out and back. The guide’s general precedence still applies ahead of it: maximise a single agent’s capabilities first, and introduce multiple agents when prompts grow many conditional branches or tools overlap enough to confuse selection.
The pattern is one vendor’s characterization of its own customer experience — the guide describes these two categories as what its work with customers highlights — rather than an independently established taxonomy, and its code examples assume that vendor’s own SDK throughout.
Related Terms
- Manager Pattern — the other category the same guide names, where one agent keeps control and calls the others as tools
- Agent Teams
- Function calling