Agent Systems: The Core Loop

The core loop is the base abstraction from which the later patterns are built.

Outline

1. Perception: observe, filter, and interpret the environment.
2. Inference: reason, plan, and choose the next action.
3. Action: call a tool, change state, or communicate with a user.
4. Feedback: feed the result back into the next iteration.
5. Explain why the loop is useful but not sufficient: the design problem is how to control state, branching, evaluation, and responsibility.

Design question

Which part of the loop is flexible, and which boundary must be deterministic and testable?

Suggested figure

00-agent-system-overview.svg in the Agent design pattern illustration folder.