From the source
Amazon's AWS published a seven-layer composable reference architecture, the Pluggable Agentic AI Framework, for taking multi-agent AI applications from prototype to production scale.
- The framework addresses a gap the post says stalls most agentic AI initiatives: prototypes show a model can do a task but are not built to survive production traffic, cost pressure and security review.
- Each of the seven layers has a distinct responsibility and a stable contract, and the services shown reflect the reference implementation rather than mandatory choices, so organizations can substitute alternatives that fit their existing infrastructure.
- The Data and Knowledge layer owns retrieval, embeddings and access to structured and unstructured data, with Amazon Bedrock Knowledge Base, Vector Engine for Amazon OpenSearch Serverless, Amazon S3 Vectors and Amazon DynamoDB as example building blocks.
- Model and Inference choices should be driven by configuration rather than hardwired into application logic, with a routing layer and a proxy such as LiteLLM providing a common interface across providers.
- Orchestration should use a single agent for straightforward tasks and a supervisor delegating specialist agents for complex cross-domain workflows, with specialists discovered from a centralized agent registry through protocols like MCP and A2A.
- Security and Guardrails should be built in from day one, treating all external content as untrusted, authenticating every request through a service like Amazon Bedrock AgentCore Identity, and granting agents temporary scoped credentials instead of persistent secrets.






