AI agents, without the hype

The word agent now covers everything from a chatbot with a search box to software that runs for hours on its own. That range makes the conversation confusing, and it makes buying decisions harder than they need to be.

A working definition helps. An agent is a language model that can call tools, look at what came back, and decide what to do next, in a loop, until a task is done or it hands the task to someone else. The interesting questions are all about that loop: which tools, which limits, and where it stops.

Autonomy is a dial

It is more useful to think about how much an agent is allowed to do without asking than whether something is an agent at all. We find five settings cover almost every real project.

• Answer: responds from your documents, takes no actions.
• Draft: prepares the email, the ticket reply or the quote for a person to send.
• Act with approval: takes actions in your systems, each one approved first.
• Act within limits: acts on its own inside clear boundaries, escalates the rest.
• Act on its own: runs end to end, reviewed after the fact.

The tools matter more than the model

Most of the work in a good agent is not the prompt. It's the tools it can call and how carefully they are shaped. A tool that returns exactly the five fields the agent needs, with clear errors, will make an average model look smart. A tool that returns a raw database dump will make the best model look careless.

“Give an agent fewer, better tools, and it will make fewer, better decisions.”

Before it meets a customer

Three things should exist before any agent goes live. A set of evaluations: real examples with known good answers, run every time the agent changes. Guardrails: the actions it may never take, enforced in code rather than in the prompt. And a cost ceiling, because a loop that goes wrong can spend money quickly.

None of this is glamorous, and all of it is the difference between a demo and a system you can leave running over a weekend.