Agents and multi-agent patterns

An agent is a model in a loop with tools, data and memory. The common multi-agent patterns call each other synchronously by default.

Now, later, never

If the caller needs the answer in this turn, call and wait. If it can come later, send a request with a correlation ID. If nobody needs to answer, publish a fact as an event.

The event bus: trigger, emit, await

An event starts an agent, an agent publishes a fact, and other agents subscribe to it, replay it or wait for it without knowing who produced it.

Long-running agents: who holds the call

When an agent runs for minutes or hours, something holds the token that resumes the caller: an AWS Step Functions task token, an AWS Lambda durable functions callback, or a completion event the agent emits itself.

Event-driven architectures

The event-driven basics this session builds on: producers and consumers that don't know about each other, with an event bus between them.

Asynchronous systems and event-driven architecture for global scale

Building next-gen applications with event-driven architectures

Event-driven architectures at scale: manage millions of events

Human in the loop with events

When the hotel agent goes over budget, the workflow sends an approval request with its token, waits without using compute, and resumes when the matching answer comes back.

When things fail

Agents run on the same at-least-once delivery as any event-driven system, so make every step safe to retry and give every multi-step booking a way to undo itself.

Seeing what agents do

A failed delivery shows up in the logs, a wrong answer from an agent does not. Trace every hop, log the bus and score agent output.

Try it yourself

AWS Step Functions orchestrates Strands agents on Amazon Bedrock AgentCore and pauses and resumes them with callback tokens.

Build, deploy, and operate agentic architectures on AWS Serverless

Building intelligent workflows with event-driven AI

Building AI agents with serverless, Strands, and MCP

Improve agent quality in production with Amazon Bedrock AgentCore Evaluations

Using multiple agents for scalable generative AI applications