This post discusses the introduction of 'Mid-Harness', a method for improving the reliability of actions executed by terminal agents in AI systems. The approach focuses on sampling and verifying candidate actions before execution, revealing significant improvements in action reliability and overall success rates through careful test-time compute allocation. The findings are supported by performance metrics from experiments using the TMAX-9B model on the TerminalBench-Lite benchmark, showcasing how action scaling can effectively enhance AI performance without changing the action generator.