Method Harness Engineering

Why the discipline around delivery — and not just around the agent — decides whether AI ships real software.

Give the same coding agent to two teams.

One ships a feature that passes code review, respects the application's architecture, runs in production, and looks like it was written by someone who knows the codebase.

The other ships a folder full of TODO, three functions that return null, a stray mock in the middle of the application, a test suite that basically confirms true == true, and a confident explanation of why everything is "done."

Same model. Same week. Completely different result.

The difference, almost always, isn't just the model's "intelligence." It's what exists around it.

Addy Osmani drew attention to this when he talked about the agent harness: prompts, tools, context, permissions, hooks, sandboxes, subagents, observability, and feedback loops that make an agent work better. Viv Trivedy took the idea further when he talked about Harness as a Service, treating the harness as a reusable layer for running agents with context, tools, filesystem, session, permissions, and observability.