FlowFn screens every public-facing agent for content that violates the platform's Acceptable Use Policy and Content Moderation Policy. The check looks at the parts of the agent you author — name, description, task / instructions, branding text (display name, greeting, placeholder), calibration summary, and the list of tools the agent is allowed to use.

When the check runs

On create and on save, whenever you edit any of the authored fields above.
On a background sweep, so agents stay in sync with policy as the model and the platform evolve.
When you publish (make the agent public or enable embed). If the cached verdict says the agent is blocked, the publish toggle is refused with a moderation reason.

What the verdict can be

Low — nothing of concern. The agent works normally.
Medium — flagged for internal admin review. You can keep editing; the agent still runs.
High — policy violation detected. The agent is taken off public surfaces automatically: is_public is set to private and the embed widget is disabled. You still see the agent in your dashboard and can edit it.
Admin blocked — an admin has manually blocked the agent. Same effect as high, but sticky — only an admin can lift it.
Admin approved — an admin has manually cleared the agent. AI re-evaluations no longer flip the verdict back.

The red banner on a blocked agent

When you open a blocked agent in the dashboard you'll see a red banner at the top of the edit page explaining what was flagged and why. The banner shows the AI's reason, any admin remark, and which policy categories matched. While the agent is in this state you cannot make it public or enable embed; the rest of the dashboard works normally.

Appealing a block — Request human review

If you think the block is incorrect, the banner has a Request human review button. Click it, write a short explanation of what the agent actually does and why it doesn't violate the platform terms (10–4000 characters), and submit. An admin reviews the appeal and you get an email with their decision:

Approved. The agent's verdict flips to admin approved — you can publish it again. The agent's content is unchanged.
Denied. The block stays in place. The admin includes a remark explaining the decision; if you address the issue and re-save, the AI will re-evaluate and may clear the block on its own.

Only one appeal can be open per agent at a time. Past appeals stay visible under the banner as historical context.

What gets exported

Appeals you have personally submitted are included in your self-service data export under agent_review_requests, including your message, the admin's decision and remark, and the timestamps — so a GDPR Article 15 request returns the full conversation.

Notes

Visitor chat content and webhook payloads are not scanned by this check — the agent owner is responsible for moderating conversations on their own agent (see the Content Moderation Policy).
The moderation check uses the same evaluator the platform uses for workflows and playgrounds; the rules and categories stay in lockstep across all three.
The feature is opt-in for existing tenants on a per-deployment basis. On deployments where it's enabled, existing agents are evaluated by the background sweep on their next pass.