Helpers
Helpers
Section titled “Helpers”For work that splits cleanly — map every caller of a function, check four repositories, read six long documents — your agent can hand a piece of the job to a helper. A helper does one narrow task with its own fresh context, reports back, and goes away.
Helpers are not extra agents you configure. They are how one agent divides a job it was already given. You see each one, what it is doing, and what it cost.
Watching a helper work
Section titled “Watching a helper work”A helper appears in the conversation as a live card the moment it starts. The header tells you what it was asked to do, which model it is using, how many steps it has taken, and the credits it has spent so far. Underneath, a step list fills in as it goes — one line per thing the helper actually did.
When it finishes, the steps collapse behind a Show steps link and the helper’s summary takes their place. If it ran out of its step budget before finishing, the card says Budget reached and the agent tells you what it got and what is still missing.
Helpers in the background
Section titled “Helpers in the background”A helper can work while your agent carries on with something else — its card is marked In the background — “I’ll have a helper map the callers while I start on the migration.” The card appears immediately and the helper also shows up on the background work bar. When it finishes, the agent is told and picks the result up; if the conversation has already ended, the result can wake the agent for one more turn.
Several helpers at once
Section titled “Several helpers at once”Your agent can also fan a job out to several helpers in parallel — up to 4 on Pro, 8 on Business and 12 on Enterprise. They appear as one group card, with a row per helper: what it’s working on, its latest milestone, and its running cost. Each row links through to that helper’s own conversation if you want the detail.
- Stop all on the group header stops every helper in it, after a confirmation.
- × on a row stops just that one.
- Stopping the run itself stops every helper it started — nothing is left billing behind your back.
- A helper waiting for a free slot reads Queued · N helpers ahead when it is waiting behind the others in its own fan-out, counting down as the line moves, and Queued · waiting for a free slot when it is waiting on your organisation’s overall limit — either way it never just looks stuck.
When the group finishes, the header totals it up — Ran 4 helpers · 152 credits — and the run receipt separates the agent’s own spend from its helpers’.
You are told the cost before it starts
Section titled “You are told the cost before it starts”Running helpers in parallel costs roughly as many credits as running the job that many times, so it is never a quiet decision:
- Before fanning out, the card states it plainly — running N helpers in parallel, roughly N× the credits of a single run.
- If your autonomy level has the agent check with you before actions that spend, the confirmation names the number of helpers, the estimated credits, and your current balance.
- If the estimate is more than your remaining balance, the agent doesn’t start them.
Answering a helper’s question
Section titled “Answering a helper’s question”Helpers working in parallel are full runs of their own, so if one needs your approval it asks — and the request appears inline, in that helper’s row, right where you are already looking. Answer it there. It is also in your approvals bell like any other request.
A single background helper works differently on purpose: it cannot ask you anything. If it hits something that needs approval, it stops and reports back that it was blocked, and the agent you are actually talking to decides whether to do that thing itself — at which point you get the request from the agent, in the conversation you are watching. This is deliberate: it keeps every approval on a run you can see and answer.
Running a Skill as a helper
Section titled “Running a Skill as a helper”A Skill can be set up to run as a helper instead of as instructions your agent follows itself. That is useful when the Skill is a self-contained job — summarise this repository, check these pages against the style guide — that shouldn’t crowd the main conversation.
In the Console’s Skill editor, turn on Run as helper agent and choose what this helper can do:
| Setting | What the helper gets |
|---|---|
| Read-only | Read and search only — it can look at things, and change nothing |
| Edit | Read, search, and edit files |
| Shell | The above, plus running commands on the agent’s sandbox |
There’s also a Steps budget — how many turns the helper may take before it has to report back. Each one costs credits, so keep it near what the Skill actually needs.
Whatever you choose, a helper never starts more helpers of its own and never asks you for anything directly. Only owners and admins can change these settings.
Keeping work apart
Section titled “Keeping work apart”When helpers work on code, each can be given its own branch so two of them never overwrite each other’s edits. A helper working that way says so on its card — working on branch task-3f9a — and, when it’s done, tells you what it left there. See Undo & branches.