Post

I'm ready to accept a definition of "agent" that I think is widely-enough agreed upon to be useful, at least in the AI engineering space: An LLM agent runs tools in a loop to achieve a goal (I've been dismissing the term as hopelessly ambiguous for years) simonwillison.net/2025/Sep/18/...

Like your definition, however, be better not to call them agents at all. Another case of anthropomorphism with AI (implying human characteristics they don't have) If something is called an agent, it must, in some sense, have agency, which, as you say, they don't.

I tried the "don't call anything agents" thing for a couple of years, but it feels counter-productive at this point - plenty of people are "building agents" now where what they mean is "LLMs achieving goals by running tools in a loop" and I want to be able to have conversations with them about it

I also like being able to say "you do realize the one thing agents don't have is agency, right?" just to stir things up!

Yeah fair point, can just start sounding unnecessarily obstructive. Classic case of semantic diffusion though

I don't understand the talk around using the word agent. We have had agent (and agentless) software for quite a while ie: terraform agent pool. That's not AI - it's a container that runs terraform. Prometheus agent, antivirus agent, logging agent... No LLM involved. "LLM agent" has no magic meaning

Over the past few years in the AI world people have said "I'm building an agent!" and then if you ask them what an agent is you find out they have a unique definition that's not the same as anyone else's This really annoyed me! I think consensus may be forming now over a reasonable definition

I care because I like to have productive conversations where everyone understands what the other person is talking about I wrote a bunch more about this in my post simonwillison.net/2025/Sep/18/...