Harness Engineering

"Agent = Model + Harness" — Everything except the model itself is the Harness.

Build systems that constrain, inform, verify, and correct AI agent behavior for any project.

Core Principles

1. Rippable: Each component works independently; safely remove when models improve
2. Incremental: Learn from failures; add configuration only when problems emerge
3. Feedback First: Success is silent; failures are verbose
4. Repo as Truth: All harness configuration is version-controlled
5. Simplicity First: Find the simplest solution; add complexity only when needed
6. Agent-Agnostic: Core constraints decouple from specific agent platforms

Four Pillars