Name the three canonical agent stopping conditions and explain why cost cap is the hardest to implement correctly

An agent runs in a loop: it reasons, calls a tool, reads the result, and repeats. Something has to tell that loop when to stop. There are three canonical **stopping conditions**. The first is the final-answer signal. The model decides it has solved the task and emits a designated phrase or a structured token. The runtime detects this by matching the output, then exits and returns the answer. The second is the step budget. The loop counts iterations and stops after a fixed number, say twenty, no matter what the model is doing. This is just an integer the runtime increments each turn. It is a hard backstop against an agent that never declares itself done. The third is the cost cap. The runtime adds up the tokens spent on every model and tool call, multiplies by the price, and stops once the running total crosses a ceiling. Cost cap is the hardest of the three. A step counter is exact and a final-answer match is a simple string check. But you cannot know a call's cost until after it runs, and each step gets more expensive as the context grows, so it is easy to overshoot the budget.

An agent loop needs explicit **stopping conditions** because the model alone cannot be trusted to halt. There are three canonical ones. The final-answer signal: the model emits a designated phrase or structured token, and the runtime detects it by pattern matching or a typed output schema. The step budget: a counter increments each iteration and the loop exits after N turns regardless of progress. The cost cap: the runtime sums token spend across model and tool calls, multiplies by per-token price, and stops when the total crosses a ceiling. Cost cap is the hardest to implement correctly. A step counter is perfectly deterministic. A final-answer signal is a simple match on model output. Cost is neither. The core problem is timing. You only learn a call's true token cost after it returns, because output length is unknown beforehand. So you must either estimate pessimistically or check after the fact and risk an overshoot. The second problem is compounding. State grows every turn, so each call sends a larger prompt than the last. Cost per step rises monotonically, and the single step that crosses the cap can be far larger than your average step, blowing past the ceiling between check points.