The Anatomy of an Agent Harness

This article defines an agent harness as everything that is not the model itself — the code, configuration, and execution logic that transforms a raw language model into a working agent. It systematically derives core harness components by working backwards from desired agent behaviors: filesystems for durable storage and collaboration, bash/code execution as a general-purpose tool, sandboxes for safe isolated environments, memory and search for continual learning, context management strategies (compaction, tool call offloading, progressive skill disclosure) to combat context rot, and planning with self-verification loops for long-horizon autonomous execution. The article also discusses the co-evolution of model training and harness design, noting that models post-trained with specific harnesses can underperform in alternative harness configurations, and that harness engineering remains a significant lever for agent performance regardless of model capability.

Source: The Anatomy of an Agent Harness