What is Agent Architecture?
Also called agent design.
Agent architecture is the arrangement of components that make an agent work: the model, the instructions, the tool layer, the memory and state stores, the control loop, and the guardrails. It describes how information flows between them and where each decision is made. Different architectures suit different tradeoffs between flexibility, cost, and predictability.
A conventional architecture has four layers. The model provides reasoning. The prompt layer assembles instructions, context, and available tools into each request. The execution layer runs tool calls and enforces permissions. The state layer stores conversation history, intermediate results, and longer-term memory. Requests flow down through these layers and results flow back up, once per step.
The most consequential choice is how much control the model holds. At one end the model chooses every step, which adapts well and fails unpredictably. At the other, code chooses and the model only fills in content, which is reliable and brittle. Real architectures place the boundary per task, giving the model freedom on interpretation and reserving code for anything with consequences.
Context assembly deserves separate attention because it is where most quality problems originate. Every step must decide what history to include, which memories to retrieve, which tool definitions to expose, and what to drop when the budget is exceeded. An agent that behaves erratically is often assembling context poorly rather than reasoning poorly, and the fix is in that layer.
Architecture is the frame that other terms fit into. The agent loop is its control flow, tool use is its action surface, memory and state are its persistence, and the system prompt is its instruction entry point. Changing an architecture usually means moving a responsibility from the model to code or the reverse, and each move trades adaptability for predictability.
Key points
- Typical layers: model, prompt assembly, tool execution, state and memory.
- The core decision is how much control the model holds.
- Poor context assembly causes more failures than weak reasoning.
- Every design trades adaptability against predictability.
In practice
Picture a scheduling agent as a stack. At the top, instructions describe how to book meetings. Below that, an assembler adds the current calendar, the last few messages, and definitions for three tools. The model returns a request to create an event. The execution layer checks the agent has permission, calls the calendar API, and writes the result back into state.