Burke Holland: The Copilot Harness Is All You Need
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Burke Holland, a GitHub technologist, argues that productivity gains come from understanding the harness, not from installing skills, MCPs, or 'weird prompts,' which he calls gimmicks despite working with AI daily.
- GitHub Copilot runs on a centralized harness across the CLI, the new Copilot app, VS Code, Visual Studio, and JetBrains, so Holland says learning the harness once transfers across tools.
- YOLO mode (/allow-all) gives agents the autonomy needed for real productivity gains, but Holland warns against running it on a local machine and recommends sandboxed environments like Codespaces or development containers instead.
- His workflow stages work in four moves: prototype first (using AI for both visual and conceptual mocks, including Mermaid diagrams), enter plan mode (/plan) with the optional /grill-me skill to surface edge cases, hand off to Autopilot for implementation, then iterate.
- Autopilot acts as an automatic orchestrator, routing exploration tasks to an Explore subagent on a small model and complex work to a General Purpose subagent on a larger model — no custom configuration required.
- Holland recommends medium-sized models like GPT 5.6 Terra or Claude Sonnet on medium reasoning, and advises sticking with one model across a given feature so prompt caching reduces token costs.
Why it matters: For developers drowning in daily AI-tool churn, Holland's case that the existing Copilot harness already covers prototyping, planning, orchestration, and iteration is a concrete cost-and-time argument: one learnable workflow replaces the chase for new skills and MCPs. The prompt-caching tip — staying on one model per feature — translates directly into lower token spend.
