Beyond the hype.

AI agent: behind the word, what are we talking about?

Very different tools get called “AI agents”. Before talking about autonomy, you need to know whether the software answers, suggests or acts in your applications.

In this article

Simon Willison, an engineer with a wide following in the AI world, collected 211 definitions of the word “agent” from fellow practitioners. No wonder two vendors sometimes mean very different things. The simplest way to find your bearings is to look at what the tool actually does.

Chatbot, copilot, agent

  • A chatbot answers your questions.
  • A copilot helps you prepare an action, but you stay in control.
  • An agent can launch actions itself, within the limits set for it.

Confusing them means expecting a chatbot to act, or fearing that a copilot will decide in your place.

What makes an agent

An AI agent brings together three things: a goal to reach, the ability to act through tools (query a database, open a ticket, send a message) and a degree of autonomy, wider or narrower. Simon Willison sums it up as a system that runs tools in a loop to achieve a goal.

The language model is only one of these elements. And a system that handles language very well does not necessarily act reliably: the fluency of an answer says nothing about the accuracy of an action.

A concrete example

A customer asks where their order is. A chatbot replies that it is “being processed”. An agent checks its actual status in the system, opens a ticket if an anomaly shows up, chases the carrier, then keeps the customer informed. It does not just talk, it chains actions toward a result.

Once you want to know how those actions are chained, and who decides on the next one, it becomes an architecture question: we covered it in workflow or AI agent, who decides what happens next.

Autonomy you can adjust

An agent can propose an action and wait for your approval, carry out low-stakes steps on its own, or act without intervention within a narrow scope.

The right degree of autonomy depends above all on the consequences of a mistake. Chasing a carrier and triggering a refund do not call for the same level of control.

Author · B.Coquisart - Project manager

Sources

GDN Journal · Beyond the hype.All articles

Get the next GDN Journal articles

News explained, in-depth analysis and real-world experience in AI, cybersecurity and software engineering.

One email per new article. Nothing else.

By subscribing, you agree to receive GDN Journal by email whenever a new article is published. Your address is processed by GDN and its email provider, Mailjet. It is never shared. Privacy policy.