Summary
Nick Saraev explains AI harnesses in plain language, covering why the software wrapped around a model often matters as much as the model itself.
He frames a bare language model as a “brain in a jar”: it predicts text, has no memory, cannot see your files, cannot act, and cannot check its own work. A harness supplies those missing abilities. Claude Code is Anthropic’s harness for the Claude family, and Codex is OpenAI’s harness for its ChatGPT models, and the video uses both as running examples while noting the ideas apply broadly. Saraev points out that the same model can perform very differently in different harnesses, which means a benchmark gain may come from harness improvements rather than a smarter underlying model.
The walkthrough covers the core components of a harness, including tool calling and schemas such as the Model Context Protocol (MCP), efficiency gains from trimming tool outputs and making software more AI-native, and verification layers that let a model test whether its code actually runs. He also shows how viewers can contribute by building their own MCP servers.
The video is aimed at non-technical viewers who want a clear mental model of how modern AI agents actually work.
📺 Source: Nick Saraev · Published October 07, 2026
🏷️ Format: Course Lesson
