Layer 2 — AI Agents · Chapter 2
Anatomy of an Agent
Most useful agents are systems, not single models. Their behavior emerges from several cooperating components.
Perception and input
The agent gathers information from text, files, sensors, software screens, databases, or messages. Good input design determines what the agent can actually know.
Planning
The agent converts a goal into a sequence of steps. Plans may be explicit, generated only when needed, or continuously revised as the environment changes.
Tool use
Tools give the agent power. They may include search, code execution, calendars, email, enterprise software, payment systems, or robotic controls. Every additional tool increases both usefulness and risk.
Memory
Short-term memory tracks the present task. Long-term memory stores preferences, prior outcomes, and reusable knowledge. Poor memory can make an agent repetitive; uncontrolled memory can create privacy and accuracy problems.
Evaluation loop
A dependable agent checks whether each action succeeded. It may inspect outputs, compare them against rules, ask another model to critique the result, or escalate to a human.
Execution boundaries
Permissions, spending limits, approved tools, audit logs, and required confirmations define the agent's operating envelope.