What is an AI agent?

AI agent, explained#

A plain chatbot maps one message to one reply. An agent runs a loop. Given a goal, the model plans a step, calls a tool to search, read or write something, looks at what came back, and decides the next step. Protocols such as MCP give agents a standard way to find and call those tools.

What makes an agent useful is the same thing that makes it risky: it can change things. A good agent design answers a few questions explicitly. What may it read? What may it write, and is each write reversible? Where must it stop and ask a person? How is what it did recorded? An agent that drafts into a review queue is very different from one that sends email on its own.

Agents fail in predictable ways. They act on stale or partial context, they repeat a failing call, and they report success when a write was quietly refused. Mitigations are equally predictable: give the agent fresh, scoped context; return refusals as readable errors; cap retries and spending; and keep a human approval step on anything that is hard to undo.

"Agent" is also used for very different scales of thing, from a single scripted workflow with a model in one step to an open-ended assistant that browses, codes and calls APIs for hours. When evaluating one, ask what it actually does in a single run and what it leaves behind.

For example, a founder asks an agent to prepare a product for launch. It reads the project's summary and existing documents, writes a requirements document, drafts positioning and a launch plan grounded in those requirements, and proposes the first few issues. It does not publish anything or start work; the founder reviews and decides.