Token cost forecasting before the agent runs, the architecture that prevented our $40K weekend

Specific because we lived this.

We had an agent that recursively called itself based on input length. One weekend, an edge case triggered the recursion. By Monday, we'd burned through a month's token budget.

The architecture we built after:

- Pre-flight cost estimator based on input size, agent complexity, expected recursion depth

- Hard budget caps per agent run, per user, per day

- Recursion depth limits enforced at the orchestration layer

- Real-time cost telemetry with alerts at 50%, 80%, 95% of cap

Result, cost surprises went from monthly to none. Lost a weekend to it, never again.

What's the production cost surprise you've had to engineer around?