On the Word ‘Agent’

1A moving target

In 2023 an agent was a loop: a model, a scratchpad, and a list of tools, calling itself until a stopping condition. In 2024 it acquired a browser. By 2025 the word had been attached to any product that performed more than one step without asking permission, and by now it is applied to workflows that a decade ago would have been called a cron job with a natural-language front end.

This is ordinary linguistic drift and it would not matter, except that procurement documents and draft regulations are now written using the term. A definition that covers both a multi-step research tool and a scheduled report generator is not going to produce coherent rules.

2What the word was doing

The useful content of the original sense was about autonomy under uncertainty: the system decides what to do next, and the decision is not enumerable in advance by the person who deployed it. That property is what makes agents hard to test, hard to audit, and interesting.

Almost everything now marketed as an agent lacks that property, which is generally a good thing for reliability and a bad thing for the clarity of the word. When the label stops distinguishing the systems with unbounded action spaces from the ones without, we lose the ability to talk about the specific risk that motivated the term.

3A suggestion

Say what the action space is. 'It can call these six functions against this account, with a spend cap and a human confirmation on writes' communicates more than any adjective, and it is the sentence a buyer, an auditor and a regulator all actually need. If a vendor cannot produce that sentence about their product, the word agent is not the thing to worry about.