One agent step — 120,000 tokens: where they go and how we fixed it

The developers of Uma Computer, an autonomous agent that performs tasks like web search, media generation, and video clipping, investigated why one user request cost 38 credits instead of the expected third of that. Billing showed 119,208 prompt tokens and only 1,018 completion tokens, meaning the model was mostly reading, not writing. The initial hypothesis was a long dialogue history, but the actual problematic dialogue had only two messages totaling 530 characters. Measurement revealed that the context per step was 13,900 tokens, and with 9 steps the total reached about 120,000 tokens; the overhead multiplied by the number of steps in the agent loop. Additionally, the provider reserves the worst-case cost with max_tokens set high, causing HTTP 402 errors even with sufficient balance. The fix involved moving preloaded context to on-demand tools, using a character budget with priorities and explicit truncation markers, and making speed expectations transparent. The article emphasizes that context should never be silently hidden from the model, and that optimizations should be verified on concrete cases, as they almost optimized for a case where history was irrelevant.