Confirm risky agent changes with Human in the Loop

Agents that work through a personal access token or the hosted MCP server can now propose a risky change and wait for a person to confirm it in GrowthBook. Routine work, such as creating drafts, keeps running without a stop, so the agent stays fast where the stakes are low.

How a held change works

With Human in the Loop enabled, GrowthBook doesn't run a covered write right away. It holds the request and answers with a link to a confirmation page. You open the link, review the change, and confirm or reject it, optionally with a note for the agent.

By default, the recommended actions wait for you. They include:

• Publishing, archiving, and deleting
• Starting, stopping, or changing live experiments and ramps
• Skipping checks, such as ignoreWarnings
• Editing settings, members, teams, environments, Projects, SDK Connections, metrics, segments, and similar configuration

Your confirmation rules can cover all actions instead, so every write waits, including creates and drafts. Reads never wait. A held request expires after an hour, and a confirmed request runs once with the token's permissions. Each decision is recorded in the audit log with the person who made it.

Confirming is not a review approval. A change that needs approval under your existing workflow still needs that approval.

Controls for the credentials agents use

GrowthBook 5.2 also adds controls for the credentials themselves:

• You can set a key or token to expire after a preset period of up to 365 days, on a date you pick, or never. Admins can require keys and personal access tokens to expire within a maximum lifetime.
• Under Customize, you pick a base role and limit it by Project or environment, so a token can do less than its owner. Plans with only the admin role offer Full access alone.
• When an integration names the teammate who requested a change, you can see what the agent did and who it acted for in revision history, reviews, and webhooks.

Getting started

Human in the Loop requires an Enterprise plan. Open Personal Access Tokens from your user menu, create a token, and choose Human-in-the-Loop under Permissions. For the hosted MCP server, choose it on the consent page when you sign in. The API keys for agents documentation covers confirmation rules, expiration policies, and who can confirm a held request.