00 Agentic AI Platform Reference Architecture
Context Diagram
Level 1 architecture of the high level components that need to be considered as part of an Agentic Solution.
Each block represents a capability that needs to reside on the solution.
| Component | Description |
|---|---|
| UI | The interaction UI for the human #UI |
| Prompt | The system prompt and user prompt that is used for instructing the LLM model. #promptengineering |
| Model | The LLM model or models that will be used in the system #models |
| Context | Content is the the term used to describe the information that is being sent back and forth to the LLM #context |
| Tokens | Text content gets converted to tokens, it is tokens that is processed by the LLM. It is token processing that you are charged for. #tokens |
| Tools | Represent systems that the LLM can interact with to complete a task. Can be either locally enabled tools that the LLM can call or external APIs called via Model Context Protocol (MCP) #Tools |
| Agent Framework | The programming framework that is used to build the Agent, examples are LangGraph/LangChain, AWS Strands, GCP A2A #agentic |
| Sandbox | An execution environment for running tools in a safe and secure environment #sandbox |
| Memory | The ability for the agent to store context across calls and communications to the LLM |
| Knowledge | A storage system for information that will not have been indexed into the LLM during Model training. Typically stored in a RAG system. Retrieval Augmented Generation |
| Guardrails | Set of controls and instructions that set LLM behaviours in place. |
| Telemetry | System for capturing and tracing the flow of data inside the Agent, including all LLM and tool calls. #telemetry |
| Governance | AI Governance structure and control set. |
| Evaluation Systems | The techniques and data used to ensure that the Agentic system behaves as expected. #Evals |
| Security | Security Harness that secures both the environment and LLM calls #llm-security |