Operating contract
Vendor Invoice Agent
High control
May decide
Needs approval
Must never do
Saved locally
Enterprise agent governance
Autonomy needs a control plane.
Define what your agent may decide, what must stop for a human, and what evidence survives after the work is done.
Set the boundaries before the agent acts.
Describe one real deployment. The charter translates autonomy, access, and impact into concrete decision rights, evidence, and escalation.
Every delegated decision needs a named owner, a visible boundary, and evidence that outlives the run.
The operating model
Control is a system, not a checkbox.
Three layers keep agent autonomy legible from first decision to final audit.
Make authority explicit.
Separate recommendations, bounded actions, approval-gated actions, and prohibited actions before deployment.
Design the stop path.
Name the owner, the threshold, and the kill switch. A stop mechanism only counts when someone can use it quickly.
Retain the why.
Record inputs, outputs, approvals, tool calls, and policy versions so review does not depend on reconstructed memory.