Pluggable Agentic AI Framework: A seven-layer composable enterprise agentic AI reference architecture on Amazon Bedrock AgentCore
The Pluggable Agentic AI Framework introduces a seven‑layer composable reference architecture for enterprise‑scale multi‑agent applications on Amazon Bedrock, defining stable contracts for each layer to enable secure, observable, and cost‑effective production deployments. The foundational Data & Knowledge layer leverages Amazon Bedrock Knowledge Base, Amazon OpenSearch Serverless vector engine, Amazon S3 Vectors, and Amazon DynamoDB for retrieval and embedding services. A model‑routing layer abstracts provider APIs via LiteLLM, allowing dynamic selection of smaller or larger models based on task complexity without code changes. This architecture gives engineers a blueprint to move from prototype agents to robust, production‑ready systems.
⚡ Key Takeaways
- Data & Knowledge layer uses Amazon Bedrock Knowledge Base, Vector Engine for Amazon OpenSearch Serverless, Amazon S3 Vectors, and Amazon DynamoDB as concrete retrieval/embedding services.
- Model routing is decoupled from agents through a LiteLLM proxy, providing a consistent request‑and‑response contract across providers.
- The seven‑layer design separates concerns (data, routing, orchestration, etc.), enabling teams to replace individual components with existing infrastructure while preserving contracts.
- Engineers can swap models or providers by updating LiteLLM configuration rather than rewriting agent logic.
- The framework emphasizes built‑in guardrails, observability, and security contracts; agents must conform to these contracts to avoid silent failures in production.
By standardizing contracts across data, model, and orchestration layers, engineers can rapidly scale agentic AI workloads on Bedrock while maintaining security, cost control, and reliability, reducing the duplicated effort of rebuilding scaffolding for each new agent.
✅ Practical Steps
- Integrate Amazon Bedrock Knowledge Base (or OpenSearch Serverless vector engine, S3 Vectors, DynamoDB) to satisfy the Data & Knowledge layer contract of “accept a query and return relevant context.”
- Insert LiteLLM as a proxy for all model calls, configuring it to route requests to the desired Bedrock or external model endpoints based on task complexity.
- Align your agent code to the framework’s stable contracts for routing, guardrails, and observability, then deploy using the reference seven‑layer architecture as a blueprint.
Want the full story? Read the original article.
Read on AWS ML Blog ↗