So I agree with @fuzzysparkle.com about this broader definition of agent too. The way I integrate LLM into my definition is that an LLM-based agent has among its capacities a language faculty, and is capable of changing its internal state and appropriately responding on the basis of human language.
I liked what I (mis?)understood to be yours, of keeping hidden state beyond what was in the conversation thread. I think taking action can mean tool use, which can extend to embodiment, but doesn't need to be anything more than continuing a conversation.
there i defined what multi-agent is, and for this i had to define what makes one agent separate from another agent. but i didnt fully define what an agent is, leaving it somewhat fluid
Comments
Physical manifestations of agents exist on a spectrum. Some demonstrate more capabilities and some less.
May be they’ll add natural language input in next release (go, clean the kitchen area) and rebrand it to Roborock Cleaning Agent.
It can receive instruction in NL, plan its course of action in NL, reveal its state in NL, and memorize in NL.
It can do some of it or all of it.