AWS shows cross-framework A2A agents spanning two accounts
A LangChain orchestrator on ECS calls a Strands agent on Bedrock AgentCore using the A2A protocol, OAuth tokens and no shared secrets.
What AWS published
On 12 November 2026, AWS published a technical how-to on building composable, interoperable multi-agent systems with the A2A protocol on AWS. Written by Amit Chaudhary for the AWS Machine Learning Blog, the post walks through a deployed reference architecture in which a LangChain orchestrator and a Strands Agents agent communicate across two separate AWS accounts. Neither side depends on the other's framework or SDK, and the two sides share no secrets.
The A2A protocol on AWS is the headline, but the more useful detail for anyone wiring agents together is the trust model around it. The protocol is described as a Linux Foundation project, with AWS among the organizations supporting it. It relies on existing authentication standards, OAuth 2.0 and OpenID Connect, for security across organizational boundaries. The full reference architecture and deployment steps are in the AWS post, which builds on an earlier entry in the same series on inter-agent communication.
Three agents, two accounts, two identity providers
The setup has three agents. On the enterprise side, an Enterprise Orchestrator and an Enterprise Context Agent both run on Amazon Elastic Container Service with AWS Fargate, inside private subnets of a VPC with no public IP addresses. On the software provider side, a Compliance SaaS Agent runs on Amazon Bedrock AgentCore in the provider's own account.
Each account holds its own Amazon Cognito user pool, so each side authenticates against its own identity provider. During onboarding the two parties exchange only public values: identity provider URLs, client IDs, scopes and endpoint URLs. After that, each side trusts tokens issued by the other side's user pool. No secrets cross the boundary.
At request time, the enterprise app takes a token from its own pool and sends a question over HTTPS to an Amazon API Gateway endpoint, then polls for an answer. API Gateway forwards the call privately through a VPC link and an internal Network Load Balancer to the orchestrator. When the orchestrator's model decides a compliance scan is needed, it fetches its own token using a client secret held in AWS Secrets Manager and sends an A2A request to the SaaS agent. That agent checks the context agent's Agent Card at the callback URL registered during onboarding, gets a token from the provider's pool, calls back for account metadata, completes the scan and returns the result.
Agent Cards replace hardcoded routing
The orchestrator uses LangChain with Amazon Bedrock, specifically Anthropic Claude Sonnet 4.6, for its reasoning. At startup it fetches the SaaS agent's Agent Card from its well-known URL and builds a LangChain tool for each skill listed there. The model then chooses which tool to call, so there is no hand-written routing logic. The post says new skills can be added by registering Agent Cards with the orchestrator, without touching agents already running.
The Context Agent is also LangChain on Bedrock, using Amazon Nova Lite, and serves internal policy rules, data classification, business criticality and escalation contacts. It is built on the official a2a-sdk, pinned to a2a-sdk[http-server]==1.2.2. The SaaS agent is built with the Strands Agents SDK.
Limits and open questions
The demo scans with synthetic data that simulates findings across accounts. The post notes that production scanning tools would connect to services such as AWS Config, AWS Security Hub or Amazon GuardDuty, but does not implement that. It gives no cost figures, no supported Regions, no latency measurements and no guidance on token revocation or credential rotation. The pinned SDK version also implies a maintenance chore for anyone copying the setup.
What this suggests for teams building agents
This suggests that framework choice is becoming less of a lock-in decision than it was, at least when A2A is the seam between systems. If the pattern holds, a small team could keep its own agent stack and still consume a vendor's agent without adopting the vendor's SDK. It likely also means more of the integration work moves into identity: which pool issues the token, which scopes it carries and who trusts whom.
What it means for building apps with AI
For people assembling apps with AI assistance, the practical lesson is to keep the boundaries explicit. Define what one part of your app is allowed to ask another, and check tokens at each hop rather than trusting a call because it came from inside your own system. If you are adding model calls to an app you are building, Draftship's guide to adding AI to your app covers that side of the work. Draftship is not affiliated with AWS, and nothing here implies Draftship supports the A2A protocol or any service described above.
Primary sources
- Build composable, interoperable multi-agent systems with A2A on AWS · AWS Machine Learning Blog, Nov 12, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.