Interrupt policy

By default, agents pause at interrupt points and wait for your input before taking action. With an interrupt policy, you can override this behavior per agent — auto-approving actions or skipping interrupts entirely.

This is useful when you orchestrate agents programmatically and want certain decisions to proceed without manual review.

Policy modes

When no policy is configured, all interrupts default to manual — no behavior change from the standard flow.

Configuring a policy

Interrupt policies are configured at the account level. Contact your account manager to configure a policy for your account.

A policy has a default mode and optional per-agent overrides:

1
{
2
  "interrupt_policy": {
3
    "default": "manual",
4
    "overrides": {
5
      "watchlist_resolution": "auto_approve"
6
    }
7
  }
8
}

In this example, all agents pause for review except the watchlist resolution agent, which auto-approves its proposed dismissals.

Auditability

Regardless of mode, the agent’s work is fully recorded:

• Actions are always executed through the same path as manual approvals. Auto-approved actions are indistinguishable from human-approved ones at the action level.
• Steps carry a resolution field on completion that indicates how the interrupt was resolved: auto_approve or skip. Manual resolutions have no explicit marker.
• Artifacts are preserved in full — the agent’s findings and recommendations are recorded even when auto-approved or skipped.

Document upload interrupts

Some agents pause to request a document upload rather than propose an action. Since there is no document to fabricate, both auto_approve and skip treat the interrupt as cancelled — the agent continues without document verification, using its initial assessment.