Relevant NewsDaily Edition
A new benchmark called Long-Horizon-Terminal-Bench introduces 46 complex terminal tasks designed to test AI agents on long-horizon planning, requiring hundreds of episodes and hours of execution rather than minutes. The benchmark uses dense intermediate rewards and partial credit grading across nine task categories including software engineering, scientific computing, and interactive games. Testing 15 frontier models reveals significant headroom, with even the strongest achieving only 15.2% pass rate on partial credit and 10.9% on perfect completion.
Why it matters: This benchmark exposes critical limitations in current AI agents' ability to handle realistic, long-horizon workflows that require iterative debugging and context management—a key gap between lab performance and real-world usability.
