Delegation & Sub-Agents

Let an agent hand off parts of a job to other agents you've built — and bring back the results.

Agents in Operator can now assign (delegate) tasks to sub-agents — other agents that work on a slice of the job and report back. The lead agent stays in charge of the conversation, picks the right specialists, briefs them, waits for their results, and combines everything into a single answer for you.

When an agent will delegate

A lead agent decides on its own to delegate when:

• The workload is too large to handle well in a single run.
• It can't complete the work itself — it lacks the right capabilities or tools for part of the job.
• A more specialized agent is clearly better suited for a sub-task than it is.
• You explicitly tell it to — e.g. "use my Invoice Reconciler to go through last month's batch".

Delegation happens automatically. You don't have to name the sub-agents or wire up any routing — the lead picks them based on their names, descriptions, and what you've asked for.

What can be a sub-agent

Delegation only ever targets your user-defined agents. A lead will never delegate to a base agent like Business Specialist, Advisor, App Builder, etc. — those stay focused on their own purpose.

This is why investing in a few well-scoped user-defined agents pays off so much: as soon as you've built a clear specialist for a recurring task, your lead agents will reach for it on their own. A user-defined agent can also delegate to a copy of itself, which is useful for fanning out the same operation across a long list.

To make a user-defined agent a good candidate for delegation:

• Give it a clear name and description so the lead knows when to pick it.
• Write focused instructions for the one thing it does well.
• Turn on the capabilities it needs (data access, email, calendar, etc.) — and pre-approve any destructive ones you don't want to confirm every time (see "Confirmations" below).

See Creating Agents for the full guide on building user-defined agents.

The trade-off: cost and speed

Splitting work across sub-agents is not free. When an agent delegates, latency, total execution time, and AI credit consumption all go up compared to a single agent doing the same work end-to-end. In general, delegation will not be faster than a single execution, and it will certainly be more expensive.

That said, there are real reasons to accept that cost:

• Context discipline. With a large, varied workload, each sub-agent focuses on its own slice and keeps its context window clean. The lead agent stays at the high level and doesn't have to clutter its own context with the details of every sub-task.
• Parallelism. When sub-tasks are independent, they can run in parallel. Even though the total credit cost is higher than a single run, the wall-clock time can be shorter.
• Role-specific assistants. Delegation makes it practical to build role assistants that coordinate specialized agents on your behalf — a Sales Assistant, an HR Assistant, a Supply Chain Assistant, and so on. Each user can primarily interact with a single role assistant and let it route the work to the right specialists behind the scenes, instead of picking the right agent themselves every time.

For small, conversational requests there's no benefit — the lead agent just answers directly.

Which agents can delegate

Delegation is an opt-in capability on a small set of base agents. You'll find it as Agent Delegation on the Capabilities page of these agents (and any user-defined agents derived from them):

• Business Specialist
• System Architect
• Partner

It is off by default. Turn it on for the agents you actually want to act as "leads".

It is not available on Advisor, Sales Executive, App Builder, Site Builder, or Pure AI — those agents are designed to stay focused on their own thing.

How to set it up

1. Build (or share in) the user-defined agents that will act as your specialists. Keep each one focused on one job.
2. Open the Capabilities page of the agent you want to use as the lead (e.g. your Business Specialist).
3. Enable Agent Delegation.
4. Save. From the next conversation onward, that agent can pick from your specialists whenever a request fits.

What you'll see in the conversation

When a lead agent delegates, the conversation shows a nested status card per sub-agent:

• Running — the sub-agent is working.
• Awaiting confirmation — the sub-agent wants to do something destructive (see below) and is waiting for you.
• Completed — the sub-agent finished and returned its result.
• Failed — something went wrong; a short human-readable reason is shown (for example, "timed out — the task ran longer than the platform allows").

While anything is still in flight, the parent conversation also shows a "Waiting for sub-agents…" indicator so it's clear the conversation is alive and just waiting on background work — not stuck.

You can expand any card to see the full transcript of that sub-agent's run, exactly as if you had opened its conversation directly. Once everything is done, the lead agent posts a final summary that combines all the sub-agents' results.

Confirmations and safety

Anything that would change real data — creating, updating or deleting records, sending emails, etc. — still requires your approval, even when it happens inside a sub-agent. Those confirmation prompts bubble up to the main conversation so you can approve or reject them in one place, without having to dive into each sub-agent.

If you've already pre-approved a capability on the sub-agent (in its Capabilities page), it won't ask again — same rule as for any other agent.

You can also cancel a sub-agent at any time from its status card if you decide its work is no longer needed. The lead agent will continue with whatever the other sub-agents produced.

Long-running and background work

Sub-agents run in the background on the platform's own schedule, not inside your browser tab. That means:

• You can close the tab or navigate away; the work keeps going.
• A delegated job can run much longer than a normal single response — if a sub-agent needs more time, the lead simply waits and resumes when results are ready.
• When everything finishes, the conversation updates with the final answer the next time you open it (or live, if you're still watching).

Limits to be aware of

• Maximum 5 sub-agents per delegation step. A lead can delegate again later in the same conversation if more work shows up.
• One level deep. Sub-agents can't themselves delegate further — for now.
• Sub-agents inherit the lead's instance connection, enterprise context, and reasoning level, so they operate against the same data and at the same level of care.
• Cost is per sub-agent. Each one consumes AI credits from the same budget as the lead, so delegation is most economical when each sub-task is genuinely worth its own run.
• Built-in agents cannot be targeted. If you want a specialist to be delegatable, it must be a user-defined agent.

Tips for getting the best results

• Name and describe your user-defined agents clearly. The lead picks specialists based on their description, so "Asset Importer — turns supplier spreadsheets into Fixed Assets" works much better than "Helper 3".
• Keep each specialist narrow. A small set of focused agents beats one mega-agent for delegation.
• Pre-approve safe destructive actions on routine specialists (e.g. "create calendar event") so batch jobs don't pause for confirmation on every item.
• Consider a role assistant. If you keep choosing between the same handful of specialists, build a single "Sales Assistant" / "HR Assistant" / "Supply Chain Assistant" with delegation turned on, and let it route to them for you.
• Watch the first few delegated runs. Expand the sub-agent cards to confirm the lead is briefing them the way you'd expect — and tweak the specialists' instructions if not.