Composable Multi‑Agent Systems with A2A on AWS
· Engineer's Notes · Cem Koyluoglu
A reference architecture demonstrates cross‑framework, cross‑boundary agent collaboration using the A2A protocol, OAuth, and Amazon Bedrock.
What happened
A reference architecture was deployed across two AWS accounts, one belonging to an enterprise and the other to a SaaS provider. The architecture connects a LangChain‑based enterprise orchestrator with a Strands Agents‑based compliance agent through the Agent‑to‑Agent (A2A) protocol. OAuth 2.0 and OpenID Connect are used for authentication, and Amazon Cognito user pools provide identity for each side. The orchestrator runs on Amazon ECS with Fargate, while the compliance agent runs on Amazon Bedrock AgentCore. Both sides exchange only public onboarding values, such as identity provider URLs, client IDs, scopes, and endpoint URLs. After onboarding, the orchestrator forwards user queries to the compliance agent when a compliance scan is required. The compliance agent retrieves context from the enterprise context agent, performs a synthetic scan, and returns results to the orchestrator, which then presents the answer to the user via an API Gateway HTTPS endpoint.
Why it matters in production
The architecture demonstrates several production‑relevant properties. First, the use of A2A eliminates the need for custom integration code between heterogeneous frameworks, reducing development effort and maintenance overhead. Second, the separation of internal tools from the external interface protects vendor secrets and allows independent evolution of internal components. Third, OAuth token exchange between the two Cognito user pools provides fine‑grained, cross‑boundary security without sharing secrets. Fourth, the system relies on managed services such as Amazon Bedrock, ECS, and API Gateway, which scale automatically and reduce operational cost. Fifth, the bidirectional pattern, where agents call each other as peers, supports dynamic workflow composition and avoids a single point of failure. The deployment also illustrates cost considerations: the enterprise account incurs charges for a NAT gateway, two Fargate tasks, an internal Network Load Balancer, and two interface VPC endpoints, while Amazon Bedrock and Bedrock AgentCore charge for usage. The architecture uses private subnets and VPC links to keep traffic internal, eliminating public IP exposure.
Engineering takeaways
- Deploy agents in separate AWS accounts and use distinct Amazon Cognito user pools to enforce trust boundaries.
- Exchange only public onboarding values; trust tokens from the other side’s pool to avoid secret sharing.
- Use the A2A protocol for all cross‑boundary communication; keep internal tool access confined to the agent’s own framework.
- Leverage Amazon Bedrock foundation models for reasoning and Amazon Nova Lite for lightweight metadata lookups.
- Store client secrets in AWS Secrets Manager and retrieve them within the orchestrator when needed.
Related work on this site
- Agentic & LLM Automation — Multi-step LLM pipelines that run unattended: model cascades with fallbacks, validation gates, scheduled automation and alerting when something breaks.
- RAG & Grounded LLM Systems — LLM features that answer from the right source instead of from memory: document-grounded assistants, transcript-grounded chat and quality gates around generated output.
- YouTube AI Summarizer — Open-source Chrome extension on the Chrome Web Store: AI summaries, key points, deep analysis, a two-host AI podcast (Gemini TTS) and transcript-grounded chat for any YouTube video, using Groq or Ollama Cloud with the user’s own key.
- Automated AI News Pipeline (this site) — The pipeline behind this site’s Tech News: scheduled GitHub Actions scrape sources, an LLM cascade on Groq rewrites and enhances articles, and layered quality gates (date integrity, language checks, instruction-leak detection, duplicate detection) decide what gets published, with Telegram alerts.
This note was drafted with AI assistance from the primary source credited on this page, and automatically checked against that source before publishing.