TerminalBench

TerminalBench is a dataset used to evaluate the performance of monitors in predicting failures in large language model (LLM) agent tasks based on terminal outcomes.

Papers using TerminalBench (5)

Sakana Fugu Technical Report2026Failure as a Process: An Anatomy of CLI Coding Agent Trajectories2026Agents Explore but Agents Ignore: LLMs Lack Environmental Curiosity2026PrefixGuard: From LLM-Agent Traces to Online Failure-Warning Monitors2026Combee: Scaling Prompt Learning for Self-Improving Language Model Agents2026