Build composable, interoperable multi-agent systems with A2A on AWS

The Agent-to-Agent (A2A) protocol enables composable, interoperable multi-agent systems across organizational boundaries. This article explores a reference architecture on AWS that demonstrates this capability. Three agents run in separate AWS accounts, one for an enterprise and one for a SaaS provider. The A2A protocol facilitates secure communication between these agents without requiring shared secrets or forcing both sides to use the same framework.

When onboarding, both agents exchange only public values like identity provider URLs, client IDs, scopes, and endpoint URLs. They then trust tokens issued by each other's user pools, eliminating the need to share any secrets. After onboarding, each request follows a step-by-step process. The enterprise app obtains an access token from its Amazon Cognito user pool and sends a question to the Enterprise Orchestrator via Amazon API Gateway.

The orchestrator, a LangChain agent built with Amazon Bedrock's Anthropic Claude Sonnet 4.6, decides if a compliance scan is required and obtains its own token from the enterprise user pool through AWS Secrets Manager. It then sends an A2A request to the compliance SaaS agent on Amazon Bedrock AgentCore runtime, which checks the enterprise context agent's Agent Card and retrieves a token from the SaaS provider's user pool.

The compliance agent calls back to the enterprise context agent for account metadata, performs its scan, and returns the answer to the orchestrator and the app.

The enterprise side consists of two agents running on Amazon Elastic Container Service (Amazon ECS) with AWS Fargate. The Enterprise Orchestrator serves as the user interface, while the LangChain framework and Amazon Bedrock provide the necessary agent tools. The SaaS side runs a Compliance SaaS Agent on Amazon Bedrock AgentCore in the SaaS provider's AWS account, built with the Strands Agents SDK. This agent performs compliance checks in AWS accounts.