AI agents acting autonomously across enterprise systems create a governance problem that abstract policy documents cannot solve. The core issue: controls placed at the agent layer are only as reliable as the agent's output is predictable, and autonomy is precisely what makes that output unpredictable. EDB's framework, outlined in their white paper 'Governing Agentic AI at Enterprise Speed,' argues that enforcement has to move to the data layer itself, enforced at query time, in the moment the agent requests access, not before or after.
The practical architecture resolves into nine controls across three categories. Enforce it: role- and attribute-based access control applied to agents as first-class principals, dynamic column masking, and declared purpose bound to agent identity at session start. See it and prove it: data classification that drives policy, session-level audit logs capturing agent identity, acting user, and declared purpose, plus lineage tracing across pipelines. Unify and harden: centralized portable policy management, encryption at rest and in transit, and consistent enforcement across on-prem, cloud, and air-gapped environments. What is new is not the mechanism but the recognition that agents need their own identity and declared purpose evaluated in the same policy path as human roles.
The white paper is worth reading in full for the 'declared purpose' concept specifically. Binding an agent's stated intent to its session identity means the access layer evaluates purpose the same way it evaluates department or role today, and the audit record proves it afterward. This is the detail that separates a governance framework from a governance story. EDB builds this on open-source Postgres, which matters for regulated industries where data sovereignty and inspectable enforcement are preconditions for production deployment, not optional features.
[READ ORIGINAL →]