Has anyone actually solved runaway agent costs? Looking for patterns beyond logging

This is exactly what we landed on too. The non-bypassable external gate is the only real answer — agent asking its own check is too easy to skip. The atomic ledger problem is real. We solved it with a per-request check against a server-side policy engine rather than a local counter, so concurrent calls can’t race past the cap. Would genuinely love to compare notes. I shipped what we built as pip install openai-agents-valta — happy to walk through the implementation if useful.