Terminal-Bench

What it is

Terminal-Bench evaluates how well AI agents complete real end-to-end tasks inside a sandboxed shell — installing dependencies, running builds, debugging failures, and driving multi-step command-line workflows. It has become one of the reference scores for agentic coding ability, cited alongside SWE-bench when comparing frontier and open-weight models, and ships a harness so teams can run their own agents against the task suite.