An approval workflow is a mechanism that pauses an AI agent when it tries to take a high-risk action and routes the decision to a human reviewer. The action does not execute until the reviewer approves it. If denied, the action is blocked and the agent is notified.
Some actions are too consequential for full automation. Sending money to a vendor, deleting production data, deploying code, or contacting a customer are all actions where a wrong decision has real costs. Approval workflows let you automate the routine while keeping humans in the loop for the decisions that matter.
escalateEscalation is triggered by policy rules:
rules:
- tool: "payment.send"
action: escalate
when:
args.amount:
gt: 100
reason: "Payments over $100 require approval"
The reason field is shown to the reviewer so they understand why the action was flagged.
For most actions, one reviewer is enough. For high-stakes actions, require multiple independent approvals:
Multi-party approval prevents a single compromised or inattentive reviewer from greenlighting a dangerous action.
Every approval request needs a timeout. If no reviewer responds within the configured window, the action should be denied by default. An action that nobody reviews should not proceed. This is the fail-closed principle applied to approval workflows.
Approval requests can be delivered through:
The channel should match the urgency and the reviewer's workflow.
From the agent's perspective, an escalated action simply takes longer. The agent receives a response saying the action is pending. Depending on your configuration, the agent can continue with other tasks while waiting, or it can pause until the decision arrives.
Well-designed agents handle denials gracefully, informing the user that the requested action was not approved and offering alternatives.
Explore more guides on AI agent safety, prompt injection, and building secure systems.
View All Guides