The core loop is: observe the current state, decide on an action, execute it, and feed the result back as a new observation. Everything else in an agent framework, from memory to planning to guardrails, is an elaboration of one of those four steps.

University of Michigan - Ann Arbor

A topic of the Building AI Agents course. Planned for 2 hours at Medium difficulty.

The Perception, Reasoning and Action Loop

The Agent Loop

An observation is whatever the program hands back to the model after an action: a tool result, an error string, a file's contents, a user reply. Observations are the agent's only channel of truth about the world, so their fidelity bounds how well the agent can possibly behave.

Observation

An agent is a system that pursues a goal by choosing its own actions in a loop, rather than executing a fixed script written by a developer. The defining property is autonomy over control flow: the model decides what happens next, and the surrounding program supplies the means to act and the results of acting.

Learn Before

Related