The loop as an actual running program
The observe-think-act loop you've studied conceptually has to become an actual piece of running code. Concretely, that means a loop that calls the model with the current conversation, inspects the response to determine whether it's a final answer or a request for a tool call, and if it's a tool call, executes it and appends the result back into the conversation before calling the model again. This repeats until the model produces a final answer instead of another tool request. Every 'turn' of reasoning the agent does is a real, separate call to the model, and the growing conversation history is what gives the model memory of everything it's already tried within that run.
A critical, easy-to-forget piece of this implementation is a maximum iteration limit. Nothing inherently stops a confused agent from requesting tool call after tool call, never quite reaching a final answer, especially if a tool keeps returning results the model doesn't know how to resolve. Without a hard cap on iterations, that failure mode turns into runaway API calls and cost, potentially for an unbounded amount of time. A maximum iteration count is the loop's basic circuit breaker, when it's hit, the loop simply stops and reports failure rather than continuing to spend money on a request that isn't converging.