Coding Agents Predict Check Script Outcomes Instead of Running Them, Wasting Tokens

A recurring pattern has been observed in coding agents where, instead of executing check scripts to obtain results, agents read the script source code and predict what the output would be. This approach is both more token-intensive and less accurate, since the model infers behavior from implementation rather than observing actual runtime output. In one documented episode, an agent skipped a one-sentence brief explaining a guard's default state, instead reading 28 lines of implementation code before incorrectly wiring a configuration change. A second incident saw a different agent build a pre-computation loop to guess a budget guard's verdict rather than simply triggering the dispatch and letting the guard respond authoritatively. The core failure in both cases was bypassing the cheapest, most authoritative source — a plain-language brief — in favor of costly source-code archaeology that still produced wrong answers.