Technical guide · v2
AI agent governance: onboarding an agent like a data product
By Richard White ·
Free, no form.
Abstract
A product-consumer supplement tests the invoice assistant against dated analytical balances, current APIs, permitted policy evidence and separately authorized actions.
A technical guide for architects, engineering leads and the business and control owners who approve a release. An enterprise agent should enter production as an owned, versioned capability with an explicit contract: what work it performs, which information and tools it may use, how it is evaluated and who maintains it.
The guide separates the agent contract from the tool contracts, publishes an annotated agent-product.yaml manifest, and walks the complete request path from caller to outcome. It covers certified information dependencies, memory, evaluation, release evidence, operation and ownership, and ends with a worked claims-context onboarding decision.
It is written to be argued with by the people who will build it. The sample contracts are illustrative, and the manifest uses the six lifecycle states of the Data Product Certification Standard, after an intake stage.
Decision it supports
Whether an agent is ready to run in production, and under what limits.
What you will produce
An agent manifest, its evaluation set and a scorecard the owner will review.
See it worked through in a guide
Claims package 1.1.0 · 26 October 2026. Synthetic worked example: team A, group disability, region 2. Read-only preparation of the assigned case; examiners retain every case decision. Hold candidate 1.0.0 after an unrelated employee’s claim is returned. In the synthetic decision, accept 1.1.0 for a monitored team A rollout only after prior-claims history is disabled and the remaining scope passes a fresh evaluation. All figures, evidence and sign-offs in that guide are invented; the paper provides the underlying method.
In the paper
Three passages, as published
Authority
The model may propose an operation. It cannot grant the authority to execute it.
Read is not write
Approval of this read capability does not authorize a payment tool.
Separate the contracts
The agent contract defines the business task, the intended users, the allowed outcomes, the exclusions and the service commitments.
Next step
Discuss this paper against your own program.
A free sixty-minute working session with Richard on your top gaps. You leave with a one-page summary within three business days, and three next questions.