LHTB 2.0 Benchmark Measures AI Agent Progress Across 3-Hour Tasks

It is a benchmark for evaluating AI agents on long-horizon terminal tasks, with 77 tasks across 13 domains that measure sustained progress through dense rewards rather than only final-state success or failure.

It preserves 46 original tasks while fixing task and evaluation-infrastructure defects, extends 31 tasks, and stretches the evaluation horizon from 90 minutes to three hours with checkpoints every 30 minutes.

They capture intermediate progress that binary evaluation misses; on 30 reconstruction tasks, binary scores averaged 0.221 versus 0.653 for dense ladders.

It is useful for teams building agents for long tasks like technical operations and analysis, but it is a new research benchmark that needs independent validation before operational adoption.