Does an agent that falls for an injected fault in a benchmark ignore available evidence?

Not necessarily. A fault-injection test can mark an agent as having fallen for an injected fault, but that only means the agent ignored a correction if the correction was actually available before the agent acted. A recent study injecting faults into tau-bench-style tasks tracked, for every run where the agent went on to adopt the fault, whether contradicting evidence showed up before or after that moment. Of 44 such runs where contradicting evidence eventually appeared somewhere in the run, only 17 arrived before the agent had already acted on the fault; the other 27 arrived too late to have changed anything.

So an adoption count conflates two different failures: an agent that saw the correction and acted on the fault anyway, and an agent that never had the chance to see it. Reading every adoption as the first kind overstates how often the agent itself is the problem. The same caution applies to attributing a regression to the wrong change: the easy story is not always the one the evidence actually supports.