reading · 7 min

Anatomy of an Agent

Objective: name the four parts every agent has and know why keeping them separable is the first engineering best practice.

Every agent, however simple, has four moving parts:

• Perception — turn the environment into an observation the policy can use.
• Policy — decide the next action from that observation (the model, or a mock).
• Action — affect the world: a tool call, an answer, a step.
• Memory — carry state across loop iterations.

In the loop you built, perception was “take the next observation”, the policy was decide / the mock LLM, the action was the printed result, and memory was implicit (none yet). Later modules expand each part: tools (action), retrieval and state (memory), planning (policy).

Best practice: keep these four as separable units with narrow interfaces. You can then test the policy without a real environment, swap perception or tools without rewriting the loop, and reason about failures one part at a time. Tangling them is the most common reason an agent codebase becomes untestable.

Next: Foundations Check