How agent work gets approved

The review surface is a queue of decisions rather than a dashboard of metrics. Each item states what an agent proposes to do, why, and what it is based on: enough to approve or correct without opening four systems.

That framing matters. A dashboard tells someone that something is wrong; a queue hands them the prepared work that fixes it.

• Accept — the action executes and is recorded
• Edit — the approver adjusts the draft or the change, then approves
• Reject — the work is discarded, and the rejection becomes evaluation signal
• Escalate — the item moves to someone with the authority to decide

Approval requirements are configured per agent and per action type. Internal CRM hygiene is commonly automated early; customer-facing communication commonly stays under review far longer. Both are your call.

Widening autonomy is a governance decision made against evidence — the record of what an agent proposed and how often a human changed it — rather than a leap of faith.

Every action, approval, edit, and rejection is logged with its actor and its evidence. That history serves three purposes: audit, evaluation, and the trust case for widening thresholds.

It is also what makes an incident tractable. When something goes wrong, the question "what did this agent do, on what basis, and who approved it" has an answer.