Multi-Agent Orchestration

Many complex production tasks exceed the scope of a single agent’s toolset, instructions, and domain scope. For example, a customer request such as “Plan a two-week trip to Japan, including flights, hotels, and a day-by-day itinerary” requires separate capabilities for flight search, hotel availability, and itinerary generation. These three distinct competencies strain a single agent’s context window, dilute its specialization, and force sequential execution even when sub-tasks are independent. The previous lesson on “Prompt Flows” showed how deterministic pipelines handle predictable, linear workflows. But when the workflow demands dynamic decomposition, where the number and type of subtasks depend on the user’s input, you need agents that collaborate.

Multi-agent orchestration is an architecture in which a coordinating agent decomposes a goal into subtasks and delegates each to a specialized worker agent. Amazon Bedrock’s multi-agent collaboration feature provides this pattern as a managed service, eliminating the need to build custom routing, context management, or failure-handling infrastructure from scratch. This lesson walks through the supervisor-worker pattern end to end, covering agent registration and routing, context passing between agents, parallel execution to reduce latency, and resilience patterns that prevent a single worker failure from collapsing the entire system.

Supervisor-worker architecture

Amazon Bedrock implements multi-agent orchestration through the supervisor-worker pattern. A supervisor agentThe top-level coordinating agent that receives the user's request, breaks it into sub-tasks, delegates each sub-task to a worker agent, and synthesizes the collected results into a final response. sits at the top of the hierarchy. It receives the user’s goal, reasons about how to decompose it, dispatches subtasks to specialized worker agentsPurpose-built agents that each handle a narrow, well-defined capability such as searching flights or querying hotel availability, then return structured results to the supervisor., and assembles their outputs into a coherent response.

Bedrock exposes two collaboration modes through the agentCollaboration configuration parameter, each suited to different orchestration needs.

• SUPERVISOR mode: The supervisor agent maintains its own system instructions and reasoning capabilities. After collecting worker outputs, it synthesizes them by combining, summarizing, or reconciling conflicting information before responding to the user. Use this mode when the final answer requires cross-worker reasoning, such as building a travel itinerary that balances flight times with hotel check-in windows.

• SUPERVISOR_ROUTER mode: The supervisor acts purely as a dispatcher. It forwards ...