What is an AI agent?

The word covers two quite different things, and conflating them is the source of most disappointment. In the loose usage, anything that does work on your behalf is an agent, including a scheduled prompt that summarises an inbox. In the precise usage, an agent is a system where the model directs its own process: it decides which tool to call, reads the result, decides what to do next, and decides when to stop. Everything that runs down a path someone else laid out is a workflow, however capable the model inside it is.

That distinction is worth holding onto because it predicts what will go wrong. A workflow fails in ways you can enumerate before you build it, because the steps are yours. An agent fails in ways you cannot enumerate, because the steps are the model's. It will occasionally take a route nobody considered, and the useful question is not whether that happens but whether the blast radius of it happening is acceptable.

The three things an agent needs are a goal it can tell it has met, tools that let it observe as well as act, and a boundary. The first is the one most often skipped. A goal stated as an outcome — the invoices are matched to the purchase orders, and the exceptions are listed — gives the system a stopping condition. A goal stated as an activity — go through the invoices — does not, and the resulting behaviour looks like either premature stopping or aimless continuation, depending on the model's disposition that day.

Tools are what make an agent different from a very long prompt. Reading a file, running a query, sending a message, calling an interface: each one is a place where the system gets contact with reality rather than with its own guesses. An agent with no read tools is working blind and will confabulate the state of the world; an agent with only read tools is safe and cannot finish anything. Most practical designs give generous read access and narrow, explicit write access, which is also the cheapest safety control available.

The boundary is what the agent may change without asking. This is a design decision, not a setting, and it is the one that determines whether the system is deployable. Reversible actions inside a sandbox can be autonomous; anything that touches money, external parties, production data or a permanent record generally should not be, at least until the system has a track record you can point at.

For most businesses the honest answer to whether they need an agent is not yet, and the reason is not caution. Agents earn their complexity on problems where the number of steps cannot be predicted. If you can draw the steps on a whiteboard, a workflow will be cheaper, faster, more reliable and easier to debug — and the agent will spend tokens rediscovering your whiteboard on every run.