Skip to main content

Subagent approvals

In an AI Team, one AI Agent can call another. When the inner AI Agent (the subagent) wants to use a tool that requires approval, you still see the approval card in your main chat — not in a hidden subagent conversation. This guide shows you what that looks like and how to manage it.

The scenario​

Let's say you've built this AI Team:

  • Lead Intake (the parent you chat with) — figures out what the user wants.
  • Website Researcher — scrapes websites, looks up company info. No approval.
  • HubSpot Writer — creates and updates CRM contacts. Approval on Create Contact.

When a user says "Add Sarah from Acme as a new lead", Lead Intake calls Website Researcher (no pause), gets company info back, then calls HubSpot Writer. HubSpot Writer calls Create Contact — and stops.

What you see in chat​

In your main chat with Lead Intake, an approval card appears, just like a regular approval. The card shows the tool name (Create Contact) and the input (Sarah's data).

You click Approve or Dismiss exactly as in a regular approval. You don't need to switch conversations. You don't need to know anything about the subagent's internal chat.

What happens after you approve​

Click Approve:

  1. Create Contact runs and writes Sarah to HubSpot.
  2. HubSpot Writer receives the result and finishes its turn with a confirmation.
  3. Lead Intake receives HubSpot Writer's reply and continues the main conversation.

This is the cascade — the whole tree wakes up from one click. See How tool approval works for the full mental model.

What happens after you dismiss​

Click Dismiss:

  1. Create Contact is marked dismissed. Nothing lands in HubSpot.
  2. HubSpot Writer receives the dismissal (plus your feedback, if any) and reacts. Typically it ends its turn with something like "I didn't create Sarah's contact because you asked me to double-check the email first."
  3. Lead Intake receives HubSpot Writer's reply and decides what to do next — usually asks you a follow-up question in the main chat.
warning

Dismissing a subagent's tool call doesn't fail the whole AI Team. The cascade completes normally; the parent gets a dismissal signal instead of a result. Your AI Team stays alive and can keep working. This is deliberate — dismissing is a decision, not a crash.

What happens deeper than two levels​

The same cascade works at any depth. A parent calls a subagent which calls another subagent which uses an approved tool — you still see the card in your main chat. It doesn't matter how deep the nesting is.

Designing AI Teams with approvals in mind​

A few patterns to save yourself noise:

  • Put approval on the destructive subagent's tool, not on the subagent call itself. Turning on approval for calling a subagent means you approve every time your parent delegates. That's usually too noisy. Approving the specific destructive tool deep inside the subagent is almost always what you want.
  • Keep read-only subagents without approvals. A Website Researcher that only scrapes websites doesn't need human review. Let it run freely.
  • Separate sensitive work into its own subagent. It's easier to reason about approvals when there's one subagent for CRM writes, one for email sends, and one for research — each with its own approval settings — than a single all-powerful subagent with a dozen approved tools.
  • Write clear subagent descriptions. When your AI Team has three subagents, the parent picks based on each subagent's description. A clear description ("HubSpot Writer: use this when the user asks to add, update, or delete a CRM contact") helps the parent route work correctly and keeps the right subagent approvals relevant.

Finding a subagent approval in the Dashboard​

The Dashboard's pending view shows subagent approvals alongside regular ones. Filter by the parent AI Agent to narrow down.

The Recent Activity feed on the Dashboard lets you follow what the AI Team did, including subagent calls and their approvals. Filter by the parent AI Agent to narrow down.

FAQ​

Can my AI Team keep running while a subagent approval is pending? No. The parent is parked in waiting_for_children until the descendant resolves. Once you act, the whole cascade wakes up and the team continues.

What if two subagents need approval at the same time? Each gets its own card in your main chat. Resolve them independently — the team resumes once all pending tasks in that turn are resolved.

Can I grant approval to a teammate? Not at the individual task level. Anyone who has access to the AI Team's team in Cubeo AI can resolve its pending tasks.