Key points
- Core thesis: Mastering the harness beats tool sprawl. Burke Holland (GitHub) argues most AI-coder productivity comes from depth of harness use, not from stacking new MCPs, skills, workflows, or tricks. Skill registries are full of low-quality "slop" because agents happily generate skills even when they are not genuinely useful.
- What a harness is: A heavily engineered runtime that unifies the model, context management, tool calls, caching, sub-agent orchestration, and cross-platform consistency. Developers do not need to rebuild this infrastructure—they need to learn how to drive it correctly.
- Prompt caching: Input tokens dominate spend, and caching saves roughly 90%. Switching models, changing the reasoning tier, or toggling tools silently breaks the cache. His separate demo ("Cash Cache Miss") shows the breakpoint clearly. Bottom line: vast majority of token spend is input tokens.
- Mermaid diagrams: Supported for visualizing prototypes and architecture sketches, more direct than prose.
- Skills:
grill-me(Matt Pocock) andpostrboard(Holland) demonstrate extension points of the harness. - MCP servers: The standard interface for extending tool reach.
- The harness is all you need (mostly) — Burke Holland
- Avoid Breaking Token Cache in GitHub Copilot — Burke Holland
- GitHub Copilot: Harness Engineering training handbook
The 8-step GitHub Copilot workflow
1. Pick a tool, any tool — CLI, the Copilot app, VS Code, Visual Studio, and JetBrains all share the same underlying harness. Beginners should start with the CLI because it has no UI learning curve. The motto: "learn it once, use it everywhere."
2. Turn on YOLO mode — Use /allow-all to grant broad command-execution autonomy. Without autonomy, agents are throttled by per-action approvals, which trains humans to stop reading what they approve. Never run YOLO on a local machine; use GitHub Codespaces or dev containers as a sandbox.
3. Start with a prototype — Generate 20 mock date-picker web components in a single HTML file for side-by-side comparison. The trick also works for non-visual tasks like API designs rendered as Mermaid diagrams. Keep the same model and reasoning tier to benefit from prompt caching.
4. Plan methodically — Switch to /plan mode without starting a new session, and enumerate edge cases (can start and end dates match? Is partial selection allowed? Clearable? Is "today" always visible? Manual input? Paste support? Date format?). Stack Matt Pocock's grill-me skill for more aggressive planning: /plan /grill-me Build a date picker web component....
5. Implement with Autopilot — GitHub Copilot will prompt to switch to Autopilot, an inner loop that forces the model to follow through on what it committed to. The harness automatically orchestrates sub-agents: the cheaper Explore agent reads the codebase, while the stronger General Purpose agent handles complex work—no custom orchestration needed.
6. Human review and iteration — This is where human taste decides quality. Holland injects design guidance through his own Postrboard CSS framework as a skill. Do not settle for "good enough" AI output.
7. Rubber-duck the result — Ask a model from a different family (for example, Claude Sonnet reviewing GPT-5.6 Terra output) for cross-review, since different models have different blind spots. Combine with Autopilot for an automated improvement loop: /autopilot rubber duck this date picker implementation. When you have the result, review it carefully and make any necessary adjustments. Repeat the rubber duck review until both you and the reviewing model agree that the only items that remain have diminishing returns.
8. Profit (commit) — Stage, commit, or keep building the next feature in the same PR. Start a fresh chat session whenever the topic drifts; chat sessions are thematic, and mixing topics breaks context control.
Side fixes Holland addresses
Why this post matters now
Two events converged in the same week:
1. Anthropic's Schema Harness paper pinned the harness as the dominant abstraction for AI agents in H2 2026, lifting ARC-AGI-3 scores from 42.83% to 98.98% with Claude Opus 4.8 + Fable 5. 2. GitHub released a production-grade harness workflow written by an engineer who uses it daily.
Both point the same direction: base-model differences are converging, so the next competitive frontier is harness design, tool orchestration, and engineered cross-model collaboration. Cursor, Claude Code, and Codex are competing on harness quality, not raw model strength. Holland's real contribution is an anti-FOMO prescription: master Copilot end-to-end to cover ~80% of software engineering work; reach for MCP, custom agents, and custom instructions only for the remaining 20%.