Why your agents still need file names in prompts

The core argument. A vocal engineering figure recently claimed that switching from frontier models like Fable or Astra to cheaper alternatives such as Opus or Soul produces no noticeable difference in results. The response is direct and emphatic: anyone who believes this is prompting too narrowly. The distinction is not about peak model intelligence but about model floor -- how rarely a model does something stupid. A model that can solve novel math problems but fails to understand 'revert' is not useful for autonomous workflows.

The real metric. The key insight is that task width matters far more than task depth. A narrow prompt -- specifying exact files and exact changes -- barely tests a model's limits. Both cheap and expensive models clear that bar. But as a task becomes wider, involving more steps, more context, and longer execution time, the probability of hitting a model's failure floor compounds. A 5% failure rate per 10 minutes becomes a 50% failure rate over 2 hours.

The data. The speaker shares personal telemetry. His median prompt went from 53 seconds to 2 minutes 20 seconds between April and the present. The 95th percentile jumped from under 7 minutes to over 16 minutes. The most dramatic shift occurred when Fable and Soul arrived, doubling long-run capability nearly month-over-month. Chat comments corroborate: runs that were 5-15 minutes became 1-4 hours. These are not outliers but patterns driven by fundamental model improvements.

Who gets this wrong. The people most likely to dismiss the value of frontier models are experienced engineering leaders who have largely stopped writing code themselves. They have outsourced the middle of the workflow -- writing code -- to agents, but still handle the front-end (spec writing) and back-end (code review) manually. They test models against the tasks they no longer do, find them adequate, and conclude cheaper models suffice. They are correctly evaluating against their own outdated workflow, not against what is now possible.

The organizational trade-off. The speaker makes the case explicit: tokens are cheaper than engineers. If a senior engineer can be three times more productive by spending their salary on tokens, that is a no-brainer, especially given that adding headcount introduces coordination overhead. The bottleneck is no longer model capability but workflow design. Teams should be optimizing for how long a model can run without human intervention and how likely the output is to work without human verification.