In this case it looks like it's "AI decision-making system that learns through reinforcement learning to handle complex tasks" - with training an algorithm to control traffic lights at many intersections in a city as their illustrative example.
Thermostat is common: an AI system that can react to its environment and continually make changes based on that ongoing feedback.
The ability to act on a loop in order to reach a predefined goal. I think of ChatGPT Code Interpreter as a good example of that one.
There's a more complex definition that requires the "agent" to have a memory and access to tools and the ability to run in a loop towards a goal and continually learn from what it's doing.
A common one among less-technical groups is an assumption that an AI agent is anything that's similar to a human travel agent - something that goes away and arranges something on your behalf.
LangChain have a "what is an AI agent?" piece - https://blog.langchain.dev/what-is-an-agent/ - that punts on the whole idea of providing a binary definition and instead talks about "agentic" as more of a spectrum. They also define an agent as something that "uses an LLM to decide the control flow of an application" - I don't fully understand what that means though.
I collected several hundred more in a bunch of Twitter replies a while back, which I've been meaning to attempt to group into categories: https://twitter.com/simonw/status/1843290729260703801 - partially scraped collection of replies here: https://gist.github.com/simonw/bdc7b894eedcfd54f0a2422ea8fea...
There's a cynical definition that I quite enjoy: LLMs that do something useful.
CrewAI is already on the radar :)