Your observation about the real failure mode being 'agent makes 10,000 correct $0.02 decisions that collectively don't make sense' is spot on. That's why per-call rate limits don't work, you need policy that evaluates context and aggregate state. We built nornr.com for exactly this: agents request a mandate before spending, policy evaluates (budget, intent, history), returns approved/queued/blocked with a signed receipt. Works with existing payment rails. Happy to compare notes on what you've learned running this in production.