Helpers

A helper is a short-lived agent that one of your agents makes for a single task: look these three suppliers up, summarise this document, check that order. It works with some of the agent's own tools, answers, and is gone. It is the quick way to split work in parallel without setting up a Team agent.

Switch it on

In the agent's setup, under Limits, turn on Can send helpers. The agent can then hand a focused task to a helper with the spawn_subtask tool. (A workflow agent runs its workflow, so it has no such switch.)

Helper model says what its helpers run on:

• Same as this agent, the default;
• Quick-step model, the workspace's Model for quick steps from Settings › Models, for cheap small work;
• or any model you can use.

What a helper may do

• Only what the agent may. A helper gets a subset of the agent's own tools, those that run without asking. It never gets a tool the agent's tool policy blocks or would ask about, and the agent is told which it held back and why: Not given to the sub-agent: run_bash (it needs a person's approval — call it yourself and they will be asked). A step's tool restriction reaches its helpers too.
• It asks nobody and sends no helpers of its own.
• It carries the agent's context marks: if the agent had read untrusted content, so has its helper.
• It is bounded in code: 10 model calls, 5 minutes when the agent waits for it (15 in the background), and at most 3 working at once in a conversation. At a limit, the tool says no and not to try again.

Waiting, or working in the background

By default the agent waits for its helpers and carries on with their answers. Several helpers sent at once work side by side.

In the app's chat, an agent can also send a helper to work in the background: the agent's turn ends and the conversation is free, the helper's line stays in the chat with Stop, and when the helper finishes its result comes back to the agent once, marked as From your helper (not from you). If the agent's turn is still running, it reads the result between its own steps; otherwise the agent takes a turn of its own and posts a follow-up. Results that finish together come back as one.

Background helpers are for the app's chat only. In a workflow step, on a messaging channel, the chat widget, a phone call or the API, a helper always works while the agent waits, so a step or a visitor's message never starts work that outlives it. A visitor's turn never sends helpers to the background.

What you see

Each helper has its own line in the chat while it works: its name, and Stop. Open it for What it was asked, Its calls (with any it was not allowed to run) and Its answer. The line is kept with the turn.

Save as agent on a helper's line opens the agent editor filled in from the helper: its brief as the instructions, its tools and its model. Nothing is saved until you press Save.

Limits and cost

On Cloud, each plan sets how many background helpers the workspace may have working at once: Free 1 · Solo 3 · Team 10 · Growth 25 · Business 50.

• A helper the agent waits for adds its minutes to the agent's turn in chat, and counts as one task on a workflow step (1 credit each; see Team agents).
• A background helper is a chat turn of its own, charged by its minutes, and the follow-up it starts is another. A wait is never charged.

Model tokens are billed as they are used. Self-hosted, a helper costs only its model tokens.

In a Team agent, helpers count against the team's Members working at once and Task budget per turn.