Summary
Cole Medin introduces and defines “harness engineering” — a term gaining traction in agentic AI circles — and breaks down what it actually means for developers building with AI coding assistants. The core idea is that the model itself (Claude, GPT, etc.) is only one layer; the wrapper built around it, what Medin calls the AI layer, is what determines real-world performance.
Medin outlines six components that make up this AI layer: global rules, skills and MCP servers, codebase search tools like LSP and knowledge graphs, hooks, and sub-agents. He argues that choosing a coding tool like Claude Code or Codex is itself a harness decision, but the higher-leverage work is customizing the AI layer on top of it. The video includes a companion GitHub repo illustrating what a mature AI layer looks like in practice.
The second half scales up to multi-agent harness engineering — stringing together multiple coding agent sessions to handle tasks too large for a single context window. Medin also covers the Google Cloud Agent CLI (the video’s sponsor), which packages deployment-ready skills for building and shipping agents via the Google Agent SDK. Whether you’re optimizing a single Claude Code session or orchestrating parallel agents across a codebase, this video provides a useful conceptual framework and concrete starting points.
📺 Source: Cole Medin · Published May 28, 2026
🏷️ Format: Deep Dive
