Sierra's agent-builder benchmark stalls at 23.9%

Sierra’s open-sourced hyper-τ-bench gives its best solo setup only a 23.9% pass rate on held-out tasks, versus 82.2% when the same class of model works with an engineer who has deep context. The benchmark report says agents often inspect fewer than 80 of roughly 1,700 files and can run 3.0x over budget, so task completion hides specification and cost failures. Teams following agent-workday economics should keep human context recovery in the loop until their own held-out score closes that gap.