The harness now enforces the policy

What the paper argues

Omnigent, the meta-harness Databricks announced, wraps every coding agent in one sandboxed session and lets policy be set at the layer above them, giving teams one control point across Claude Code, Codex and any framework they run. Prompt-only constraints were never governance. They were suggestions the model could talk itself out of, so moving the rule to where the action passes through is a real step.

The harness can prove the policy fired. It cannot prove the policy was doing any work. Level 4 is a control you can document and show ran; Level 5 is a control that demonstrably changes behaviour at the point of action. Omnigent delivers a cleaner Level 4 than the industry has had, and the distance to Level 5 is still the team’s to close.

Key points

• Each of the three policy surfaces, cost budgets, allow, deny and approve rules, and approval steps, produces an artefact you can point to. None of them, on its own, tells you whether the policy changed what shipped.
• A control that never binds, a deny rule nothing trips and an approval no one reads all log perfectly, and all three are theatre. An approval cleared in seven seconds is a log line, not a judgement.
• The Cognitive Delivery Index does not enforce policy. It reads what the enforcement layer produces and renders an independent judgement on whether the governance is substantive or theatre, scored the same way across every platform an estate runs.
• The judgement has to come from outside. A platform grading the governance it enforces is marking its own homework, and for the AI Act, DORA and ISO 42001 that is exactly the conflict an auditor discounts.
• A harness improves the raw material for several of the six dimensions and scores none of them. A constraint that fires is Constraint Fidelity evidence only if it demonstrably held; an approval that logs is Review Substance only if the review was real.
• For each policy a team has defined, find the last time it actually blocked, paused or reshaped an action. If you cannot, the control is Level 4: write down what Level 5 evidence would look like, and instrument for it before adding another rule.