A model answering from memory is guessing about the present. An agent is the same model given tools and a loop.
The loop is almost embarrassingly simple. Think: what do I need next. Act: call a tool. Observe: read what came back. Repeat until done. Each observation goes back into the context, so the next decision is made with evidence rather than assumption.
Two variations cover most real systems. Plan-and-execute writes the whole plan upfront and then runs it — cheaper, more predictable, and helpless the moment step two reveals the plan was wrong. Reflection adds a critique step before committing: draft, review, improve. It catches real errors and costs another round trip.
The failure mode is always the same: a loop that never converges. Cap the steps, set a timeout, and treat tool errors and empty results as outcomes the loop has to handle.
Intelligence in a system rarely comes from one brilliant decision. It comes from being able to look, and revise.