AI Agent Approvals
Human oversight
where it matters.
You can’t review every agent action — but you can make sure the ones that matter pass through a human. MUTX lets you define approval gates, track authorization decisions, and keep a complete record of what required oversight and what didn’t.
Approval properties
Approvals that
actually block.
Most approval systems are advisory — they suggest a human review but don’t actually prevent the agent from proceeding. MUTX approvals are control plane gates — the agent waits until the approval is recorded, and the approval record travels with the trace.
Approval workflows
Define which operations require a human to sign off before the agent proceeds. Approval gates are control plane records — not a Slack message that may or may not get answered.
Authorization tracking
Who approved what, when, and why — recorded in the audit log with the full execution context. Not just “approved by jane@” but the full chain from request to approval to outcome.
Escalation paths
When an approval is pending, the right operator gets notified through the control plane — not through whatever notification channel happens to be monitored that day.
Automatic vs. approved
Distinguish between what the agent did autonomously and what required human sign-off. This distinction is a first-class property in MUTX — visible in traces, cost attribution, and audit logs.
Connected surfaces
Approvals touch
everything in the plane.
When approvals are part of the control plane, they integrate cleanly with audit logs, governance policies, and monitoring traces. The approval record is attached to the trace, recorded in the audit log, and visible in the governance surface — not siloed in a separate workflow tool.
Governance
Approval requirements are governance policies. Who can approve what, and under what conditions — defined in the governance model and enforced by the approval gate.
Audit Logs
Every approval decision — who approved, what they approved, the full context — is in the audit log. Compliance reviewers see the complete chain from request to outcome.
Monitoring
The distinction between automated and approved actions is visible in every trace. When you’re investigating an incident, you see immediately which actions required human sign-off and which didn’t.
Guardrails
A guardrail violation can be configured to require approval before the agent can proceed. The response to a safety boundary isn’t just a log entry — it’s a workflow that brings in a human operator.
Get started
Define an approval gate
and watch it block.
Download the Mac app, define an approval workflow for a high-stakes agent operation, and trigger it. See the gate block the operation, notify the right operator, and record the approval — on the record, in the control plane.