Terminal-Bench 2

'Terminal-Bench 2' is a dataset used to evaluate the performance and capabilities of AI agents and large language models by providing a collection of benchmark tasks that can be systematically audited for issues such as ambiguous task design and execution environment conflicts.

Papers using Terminal-Bench 2 (3)

Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses2026 · 1 citesShepherd: Enabling Programmable Meta-Agents via Reversible Agentic Execution Traces2026Automated Benchmark Auditing for AI Agents and Large Language Models2026