Abstract
You wrote the AGENTS.md. You documented the error handling conventions, the package structure, the interface patterns. And then the agent ignored you.
Instruction files are inferential guides: the agent interprets them probabilistically, weighs them against everything in its context window, and makes a judgment call. That call is wrong often enough to matter — degrading further as tasks grow complex. A better AGENTS.md just moves the ceiling.
The fix is understanding what a harness is. Agent = Model + Harness: the harness combines guides - what steers before the agent acts - and sensors - what observes and corrects after, both on a spectrum from inferential to computational.
We'll explore best practices: wiring agent hooks to enforce linting on every file write, writing custom analyzers that turn conventions into compiler errors, structuring table-driven tests as specifications the agent must satisfy, and more.
Your AGENTS.md is a lie. Your harness isn't.
