The provider bill tells you the total and nothing else. This is a forensics question, and the answer is almost never a pricey model. It is that the transcript grows every turn, so a forty step run costs six times what everyone estimated.
Updated Aug 2026 · Grounded in real Forward Deployed Engineer interview loops and written to a senior-engineer editorial bar.
The provider bill tells you the total and nothing else. This is a forensics question, and the answer is almost never a pricey model. It is that the transcript grows every turn, so a forty step run costs six times what everyone estimated.
Two things separate strong answers. The first is refusing to name a fix before naming a step: candidates who open with 'switch to a smaller model' have skipped the diagnosis and usually cut quality to save the wrong 15%. The second is doing the arithmetic out loud. An agent loop resends its whole transcript every turn, so input tokens grow quadratically in turn count, and once you see that you also see why the cheapest fix is a cap on loop length rather than a cheaper model.
No comments yet — be the first to share your approach.
