AI agent governance is the set of policies, controls, approvals, and evidence mechanisms used to determine what AI agents may access, which tools they may invoke, and which actions they may perform.
The object being governed is delegated authority: the access and actions entrusted to software for a particular task. That delegation needs boundaries even when the agent can choose its own intermediate steps.
Why it matters
The functions available through a tool can exceed the authority intended for a task. An agent may be able to request a change that the organization has never permitted it to make. Governance connects the assigned purpose to the permissions and decisions that constrain its operations.
OWASP's guidance on Excessive Agency recommends limiting tool functionality and permissions, with authorization enforced in downstream systems. A model's recommendation about what to do cannot replace that authorization.
How the parts differ
Permissions set access limits for an identity or tool. They help bound what the workflow can reach, but a broadly available permission does not establish that every use of it fits the task's approved purpose.
Policy decisions evaluate a request under the applicable rules and context. They can permit it, deny it, or require approval. Where human review is required, the approval must apply to the operation being considered, with an approver authorized for that scope.
Controls give those decisions an effect. A preventive control must be able to constrain the relevant operation before it occurs. The result depends on the actual integration, including how the executor handles missing or failed decisions.
Evidence supports reviewing what was requested, decided, and observed in execution. These records answer different questions. A recorded permission does not establish that an operation completed, and a denial record alone does not prove that the executor stopped it.
What the term does not guarantee
Agent governance does not require a human to approve every action. Policy can permit a bounded class of operations while reserving exceptions for review. Nor does the label establish that a control covers every agent, tool, or execution path. Coverage needs to be established for the system being governed.
Related concepts
- AI governance covers the broader organizational decisions about developing, deploying, and using AI, including model suitability and accountability.
- Agent observability provides information about agent activity. Its records can support governance; the existence of those records does not establish an effective execution control.
- An AI control plane coordinates policy configuration and decisions; its connection to enforcement must be explicit.
The AI agent governance pillar explains how to apply these distinctions to a workflow's tools, approvals, execution path, and evidence.