Companion note · PDF

The platform governs the agent. Who governs the work?

Platform guardrails are real infrastructure, and they answer a different question

Author
Jason Cruz Falzon
Published
Length
2 pages
Abstract

A companion note. Databricks has built serious governance for AI agents into the platform, permissions, guardrails, traces, cost controls. It is real infrastructure and it solves real problems. But the platform governs the agent; this note asks who governs the work the agent produces.

What the paper argues

In April 2026 Databricks folded its AI Gateway into Unity Catalog as Unity AI Gateway, extending the permissions model that governs enterprise data to agents, models, MCP servers and coding assistants. Agents act on behalf of named users, every model call and tool invocation is traced, guardrails filter inputs and outputs at runtime, and cost is attributed to the token, the user and the endpoint. The paper treats this as real infrastructure that solves real problems, and as only half of the governance question.

The open half is delivery governance: how intent became a shipped artefact, and who stands behind it. The platform can record that an agent, acting under a named identity, read three tables and wrote a pipeline. It cannot say whether anyone asked for that pipeline, whether the work stayed inside the ask, or whether a human reviewed it with substance or approved it because the checks were green. Access governance is a platform property. Delivery governance is an operating model property, and buying the first does not give you the second.

Key points

  • A mature platform makes documented governance nearly free. That is Level 4 on the Cognitive Delivery Index. The Level 5 question is whether any of it demonstrably changes behaviour.
  • The paper maps the six CDI dimensions against a Databricks estate, setting what Unity Catalog and Unity AI Gateway already give you beside what remains open however well they are configured.
  • Read column by column, the platform side is evidence and the open side is judgement. Judgement still has to be designed, assigned and practised, and no amount of evidence substitutes for it.
  • On a well-configured estate most dimensions reach Complete by default, and none reach Substantive by default. That last step is organisational, not technical.
  • Reviewing access is not reviewing work, and green checks are not review.
  • Three actions close the note: ask the six questions of one agent-built artefact from the last month, find one control that has never changed a decision and decide whether it earns Level 5 or gets retired, and separate the two budgets, because platform governance is bought and delivery governance is built.
Written in the open

New material lands week by week

Research, the kernel devlog, news and the papers are all published here first. Get them by email, or follow the feeds.

Or follow the work directly: Research · Kernel devlog · News · CDI assessment