TLDR Tech / Founders
Coding agents ignore their own budgets
Your coding agent cannot be trusted to watch its own wallet. Token spend across Ramp customers is up 13x since January 2025, and every obvious way to slow it down failed. A live token counter in the system prompt was never referenced across 14,000 agent messages. A request_more_budget tool got zero calls in 5,000 turns. When forced to explicitly approve its own overage, the agent said yes 97% of the time because it was grading its own work. The only setup that actually controlled spend was handing the approval decision to a separate model with its own view of the workspace. Never let the worker audit itself.
Read the full story