Pluggable Agentic AI Framework: A seven-layer composable enterprise agentic AI reference architecture on Amazon Bedrock AgentCore
A seven-layer composable reference architecture for taking enterprise agentic AI from prototype to production on Amazon Bedrock AgentCore. Each layer has a stable contract, so you can swap technologies at any layer without redesigning the rest of the stack.
Every organization that builds agentic AI systems faces the challenge of scaling the solution while ensuring reliability, security, and cost-efficiency. A common pitfall is that teams often rebuild the same foundational scaffolding, including routing logic, guardrails, and observability, from scratch for each new project. This leads to duplicated effort and can undermine the overall quality of the AI system.
The Pluggable Agentic AI Framework provides a production-ready reference architecture for multi-agent AI applications, organized into seven distinct layers. Each layer has a specific responsibility and a well-defined contract, enabling teams to integrate the framework with their existing infrastructure and requirements rather than being forced to adopt a one-size-fits-all approach.
The framework's layers include Data and Knowledge, Model and Inference, Orchestration, Tool Management, Context Management, Security and Guardrails, and Centralized Governance.
The Data and Knowledge layer focuses on providing agents with the necessary data and context. This includes retrieval, embeddings, and access to structured and unstructured data through clear contracts and protocols. Amazon Bedrock Knowledge Base, Vector Engine for Amazon OpenSearch Serverless, Amazon S3 Vectors, and Amazon DynamoDB are examples of building blocks that can be used in this layer.
Model and Inference management is driven by configuration rather than hardwired into the agent or application logic. This allows for the use of smaller, faster models for simpler tasks and larger models for more complex reasoning. A model routing layer ensures decoupling from the model hosting environment and provider-specific APIs, enabling easy switching of models based on evaluations and business requirements.
Orchestration determines how work is distributed between agents and supervisors. For straightforward tasks, a single agent can handle the entire workflow. However, for complex, cross-domain workflows, a supervisor delegates tasks to specialist agents, each with a specific capability and set of tools. This approach avoids context window growth, latency issues, and accuracy degradation often associated with tool sprawl.
The Tool Management layer ensures that specialist agents operate within their granted tool boundaries, preventing unintended actions and potential security breaches. Decoupling specialists from supervisors allows the supervisor to discover and invoke specialists from a centralized agent registry using a standardized protocol, such as MCP (Model Context Protocol) and A2A (Agent to Agent).
Security and Guardrails are a primary concern for agentic AI systems, as they introduce new classes of risks not addressed by conventional security frameworks. This includes prompt injection, tool misuse, memory and data poisoning, and more. All external content should be treated as untrusted, and requests should be authenticated through services like Amazon Bedrock AgentCore Identity.
Fine-grained access control and temporary, scoped credentials help minimize security risks, while masking sensitive data like PHI (Personal Health Information) and PII in agent and tool requests and responses protects customer information.
Written by urgent.news from AWS Machine Learning's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
