Trust levels & approvals

Your agent can do a LOT of things — send email, post to LinkedIn, spend credits, move files around, schedule meetings. The question isn’t what it CAN do. The question is what you want it to do without asking first.

That’s what trust levels are for. There are three of them, and you can set them per action type.

The three modes

🛡️ Require approval (safest)

Your agent does nothing without running it past you first. It can propose an email draft, draft a social post, suggest a cron job, draft a calendar invite — but it will never hit send, publish, or save until you click Approve in your approval queue.

Use this when:

• You’re new to the platform and want to see what your agent is actually doing before trusting it
• The action is high-stakes (sending email to a customer, posting publicly, deleting files)
• You’re the only person who should ever send email from your address, regardless of how good the draft is

Trade-off: Nothing happens until you look at your approval queue. If you’re busy, things pile up.

🔔 Auto + notify (balanced)

Your agent takes the action immediately, AND sends you a push notification so you know what just happened. You can undo most actions if you see the notification and don’t like what it did.

Use this when:

• You trust your agent to make reasonable decisions but want visibility
• You check your phone frequently enough to catch anything wrong
• The action is reversible (moving an email to a folder, creating a draft, setting a reminder)

Trade-off: Stuff happens fast. If you don’t look at the notifications, you won’t know what your agent did until later.

⚡ Full auto (fastest)

Your agent just does the thing. No approval, no notification, no ceremony. You check in at the end of the day and see what happened.

Use this when:

• You’ve used the platform for a while and trust your agent’s judgment
• The action is low-stakes and reversible (labeling emails, creating draft files, running research)
• You want maximum agent autonomy

Trade-off: You have the least visibility. Your agent is essentially running your assistant on autopilot.

Configuring trust per action type

Go to Settings → Agent → Agent Trust & Approvals in the dashboard. You’ll see a list of every action type your agent can take, grouped by category:

Email actions

• Read & summarize inbox (default: full auto — it’s just reading)
• Draft replies (default: full auto — drafts aren’t sent)
• Send email (default: require approval)
• Delete email (default: require approval)
• Apply labels / move to folders (default: auto + notify)

Social media

• Draft posts (default: full auto)
• Publish posts (default: require approval)
• Comment or reply (default: require approval)

Calendar

• Read calendar (default: full auto)
• Draft event invites (default: full auto)
• Create calendar events (default: auto + notify)
• Accept/decline invitations (default: require approval)

Workspace & files

• Create workspace files (default: full auto)
• Read files (default: full auto)
• Delete files (default: require approval)

Scheduled tasks (cron)

• Create a new cron job (default: require approval)
• Pause/resume a cron job (default: auto + notify)
• Delete a cron job (default: require approval)

Sub-agents

• Spawn sub-agents for research (default: full auto)

Payments & purchases

• Make a purchase (default: require approval — always)

Other actions

• Change account settings (default: require approval)
• Anything else sensitive (catch-all for uncategorized sensitive actions — default: require approval)

How to read the approval queue

When your agent wants to do something that requires approval, it goes into your Approvals tab in the dashboard. You’ll also get a push notification (and an email, if you set that up). Each pending approval shows:

• What your agent wants to do — “Send email to sarah@acme.com with subject ‘Re: Q2 report’”
• Why — the full conversation context that led to this action
• The exact content — the email body, the social post text, the file contents
• Approve or Reject buttons

Approving sends the action immediately. Rejecting discards it and tells your agent why (you can leave a note).

A recommended starting configuration

If you’re new, this is our suggested baseline. It minimizes risk while letting your agent be genuinely useful:

• Everything “read-only” and “draft” → Full auto
• Anything that touches someone else (send email, publish post, reply to a comment) → Require approval
• Everything in between (labels, folders, calendar events) → Auto + notify

You can tighten or loosen it over time as you get comfortable with what your agent does.

Changing trust on the fly

You don’t have to lock in your settings forever. Two useful patterns:

“Just for this one thing” — When your agent proposes an action that would normally require approval, you can approve it AND set “Remember this for next time” to upgrade that specific action type to auto.

“Slow down” — If you want your agent to stop acting autonomously for a bit (maybe you’re going on vacation and want to review everything when you get back), go to Settings → Agent → Agent Trust & Approvals and click Pause automation. Everything reverts to “require approval” until you unpause.

What happens if you never approve things?

Approvals don’t expire by default — they wait for you. But if you’re swamped and approvals pile up past 100, your agent will start flagging “your approval queue is getting long” in its morning briefings and slowing down actions that depend on the stuck items.

If the queue hits 500, your agent will stop generating new actions that need approval and tell you “I’ve stopped because you haven’t been clearing approvals. Let me know when you’re ready.”

Start cautious, loosen over time

Most users start with very cautious settings for the first week, notice that their agent’s proposals are consistently good, and gradually flip more things to auto. You can always tighten back up if something goes wrong.