You said a deterministic function deciding the next ste...

What makes something an agent is whether the policy — the thing deciding the next action — was authored in advance or interpreted at runtime, not whether the decision happens to be dynamic.

A rule like “if confidence is above a threshold, act; otherwise search again” is state-dependent and looks dynamic, but it’s still a workflow, because a person wrote that rule in advance. The branching was fully specified before the system ever ran.

An agent is better described as a function of state and observation where the function itself has to interpret the situation at runtime — usually because the input is open-ended, the relevant considerations aren’t enumerable in advance, or the right action genuinely depends on judgment rather than a clean threshold.

The useful reframe this leads to: even a bounded, tool-limited agent that’s fully deterministic at zero temperature is still an agent, because what makes it agentic is that the policy is a runtime interpreter, not that its output is unpredictable. Predictability and agency are different axes; conflating them is the trap.