May 18th, 2026

Agentic Harness Engineering: Observability-Driven Evolution of Coding Agents

Agentic Harness Engineering (AHE), an automated framework designed to optimise the external components—such as prompts, tools, and middleware—that support coding agents. While most development focuses on improving base models, this research highlights that the harness surrounding the model is a critical, yet often neglected, lever for performance. By using a closed-loop system, AHE replaces manual engineering with an autonomous evolution agent that identifies failure patterns and implements precise, file-level adjustments. The system relies on three observability pillars to ensure stable growth: it decouples harness components into editable files, distils complex execution logs into structured evidence, and requires every edit to be a falsifiable contract verified by subsequent results. Empirically, this method significantly increased success rates on the Terminal-Bench 2 benchmark, outperforming both human-designed systems and existing automated baselines. Furthermore, the evolved harnesses demonstrated strong transferability, improving the performance of diverse model families and unseen tasks without further tuning. Ultimately, the sources position AHE as a practical pathway for keeping agent infrastructure advancing at the same pace as rapidly evolving artificial intelligence models.