Harness Engineering: Key Principles

Five principles hold the whole thing up, and they read like hard-won lessons more than theory:

• Explicit runtime resources: expose what the agent is spending, do not leave it implicit.
• Traceable mediation: record every selection and decision the harness makes.
• Requirement-level verification: bind “done” to a deterministic check, never to the model’s own say-so.
• Attribution before recovery: diagnose the failure before you re-edit anything.
• Maintenance and entropy awareness: track the residue and the slow degradation a long run leaves behind.

If I had to compress it further: order first, cleverness later, and system first, model second. Most of the failures I have watched come from breaking one of these quietly, usually the third one, letting the model declare victory instead of proving it.