“The most effective current applications for long-horizon agents are tasks that produce a "first draft" for human review, such as coding, research, and report generation, due to their lack of perfect reliability.”