Holland: Master Copilot Harness, Skip the AI Tool Hype
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Burke Holland argues that understanding GitHub Copilot's harness matters more than installing new skills, prompts, or chasing every new AI tool release, calling most of that churn 'gimmicks'
- GitHub Copilot runs on the same harness across the CLI, the new Copilot app, VS Code, Visual Studio, and JetBrains, so learning the workflow once transfers to every surface
- Holland recommends YOLO mode (/allow-all) so agents execute autonomously without approval-button fatigue, but pairs it with sandboxed environments like GitHub Codespaces because 'bad things happen to good people' on a local machine
- Plan mode combined with Matt Pocock's 'grill-me' skill forces the model to surface edge cases and clarifying questions—Holland insists developers must engage with the output rather than blindly accept it
- GitHub Copilot's autopilot automatically orchestrates subagents like Explore (small model for codebase reads) and General Purpose (larger model for complex actions) without requiring any custom configuration
- Holland recommends medium-sized models such as GPT 5.6 Terra or Claude Sonnet on medium reasoning, and warns against switching models mid-task so prompt caching can cut token costs
Why it matters: Individual developers gain by redirecting time from tool-hopping into one harness, but Holland's prescription to run agents in autonomous 'YOLO' mode shifts the burden to sandbox discipline—he explicitly recommends Codespaces or dev containers over local execution, framing safety as the non-negotiable cost of that autonomy.
