Approvals

AIgency layers its own permission policy on top of each agent's normal permission mode, so you can control how much a run is allowed to do without supervision.

Run-level permission mode

Every run has a permission mode:

• ask — tasks pause for your approval at points where an agent would normally need one.
• allow-approved — previously approved actions proceed automatically.
• bypass — tasks proceed without approval prompts.

Choosing bypass requires a one-time acknowledgment of a warning. The mode can be set per run, or inherited from an AIgency project's saved default.

Per-agent override

Independently of the run-level mode, each agent assignment can override behavior with inherit, trusted, restricted, or bypass — letting you, for example, run a well-tested agent with fewer interruptions while keeping a newer one on a tighter leash within the same run.

Final sign-off

Separate from per-task approvals during execution, a run has a final human sign-off step at completion — you explicitly decide whether the finished run is accepted before it's considered done.

Related

• Permissions & Safety → The approval system covers how an individual agent's own permission modes work outside of AIgency.
• Execution — the waiting_approval and waiting_interaction task states this system produces.