This article defines and analyzes the architecture of AI agent harnesses—the non-model components (filesystems, tools, sandboxes, memory systems, orchestration logic) that make LLMs functionally useful as autonomous agents. It derives harness design patterns by working backward from desired agent behaviors.
This article argues that filesystems are emerging as the primary interface for AI agents to manage persistent context and capabilities, with file formats (like CLAUDE.md, SKILL.md) replacing APIs and plugin systems as the de facto standard for agent interoperability. The author contends that the bottleneck for AI agents has shifted from model capability to context persistence, making filesystem-based architecture more practical than traditional database or API-driven approaches.