Our token costs tripled over a few weeks. We only had a few agents running, and nothing was broken, but costs went 3x what we expected. We looked at the traces, and this is what we found: • Agents were reading more files • They were loading more context • They were creating more detailed plans • They were verifying their work more frequently • They were troubleshooting mistakes • They were retrying failed steps Nothing here is broken. This is exactly what we wanted. But, as models improve,…