What is an AI agent?

An AI agent is an LLM, a harness, and a set of tools arranged in a loop, with the model rather than a fixed script deciding the next step at every turn. Given a goal, the agent reasons about the situation, picks a tool call, observes the result, and repeats until the goal is met or a limit stops it.

This places agents at one end of a spectrum. A plain chat request is one model call. A workflow chains model calls along paths fixed in code; Anthropic's building-agents guide recommends starting there, because predictable steps are easier to test and cheaper to run. An agent is what you reach for when the steps cannot be known in advance: open-ended coding tasks, research, or operating software.

The pieces in this series are the whole story. Tokens set the budget, chat defines the request, the LLM is the engine, the harness runs the loop, and tools are the hands. What makes an agent an agent is handing control of the loop to the model.

That power has a price: errors compound across iterations. Bound the loop with step limits, verify results before acting on them, and keep humans in the loop for irreversible actions.