# Subagents

> Helpers the agent launches to work in parallel, and how to create your own types.

Chats · Updated: 2026-11-10 · version 2.0.261 · https://nobreiniers.com/en/help/subagents/

For big jobs, the agent launches subagents: helpers that each do one part at the same time as others. You see them in the chat's «Subagents» panel.

For a very big coding job that is split among a planner, several workers at once and a judge who checks it, see [Long coding jobs](https://nobreiniers.com/en/help/long-work/).

## See the subagents

1. While they work, «Subagents» appears below the conversation.
2. Each one shows what it does and how it is going.

The app's own helpers, such as the one that searches your project and the one that works in the browser, use the chat's model and reason at the effort you chose for it, never less. A type of your own reasons at the «Effort» you gave it.

Each helper works on its own. One that keeps working in the background keeps its browser tabs even after the turn that called it has ended, and starts on the page its caller had open. The same goes for a task an agent leaves running for another chat: it works in its own tabs, never in those of the chat that started it.

When you press «Stop» in a chat, everything it was running stops too: its helpers, also those working in the background, and the commands it left running in the background, so nothing keeps spending after you stop it. The note that appears afterwards says what stopped and what goes on (for example a task it handed to another Nobreinier).

## How many work at once

In «Settings» (⌘,) › «Agents and helpers», «Helpers at once» says how many helpers of one chat work at the same time; the rest wait their turn and start by themselves. Type the number your work needs, from 1: there is no top. Fewer use less of your Mac.

## How deep helpers go

A helper can launch helpers of its own. In «Settings» (⌘,) › «Agents and helpers» › «Advanced», «How deep helpers go» says how many levels they may go down: from 1 to 8, and 4 if you do not change it. Fewer levels spend less; more let a big job split further.

That number holds every way a helper is launched. When a helper is already at the last level, it is told what the limit is and where to change it, and goes on with its own tools.

## Change how the helpers look

Each helper shows as a character by its role, such as the researcher, the guardian or the engineer, each with its accessory. You can change how they look.

1. Open «Settings» (⌘,) › «Agents and helpers» › «Helpers».
2. The first row is the «Default helper»: what you pick there goes for every helper. Below it there is one row per role, with how it looks now.
3. Press the «Change character» icon on a row. Pick a character from the catalog and press «Choose», or pick the Nobreinier figure with an accessory below it. It is saved at once.
4. What you pick for a role goes before the default. A row's «Reset» icon takes it back to the default, and the one at the top takes every helper back to how the app draws it.

Helpers already at work change at once. A helper type that brings its own character keeps it. You can also ask Breini, for example: «make the helpers look like Abedul».

## When a helper reaches its step limit

An agent can give a helper a limit of steps, and a type that comes from a file (a project's or a plugin's) can carry one. The form under «Kinds of helper» has no field for it.

A helper that reaches its limit no longer stops with a fixed notice and the work done lost. It makes one last pass, with no tools, and writes its answer from what it had gathered: what it found and did, with its evidence, and plainly what is still missing, undone or not verified. The agent that launched it receives that answer marked as cut short by the limit, so it knows to check what was left out.

If that last pass writes nothing, the helper ends with a notice that gives its last words, said to be unfinished and unchecked. Its work is kept, and the agent can continue the helper from where it stopped.

The same holds for any turn that has a step limit: it closes with an answer written from what it gathered, and only when that answer cannot be written does it say that it stopped after that many tool steps, the safety limit, and that the work done is kept. Write “continue” and it goes on from there.

## A helper's work checked against what was asked

When an agent hands a helper a task with conditions to meet (for example «the tests pass» or «the file exists»), the app checks each condition against what the helper really did: the commands it ran, where it ran them, how they ended and what their output said. A test run counts only when it started in the helper's own folder and tests really ran and passed; a run that ran no test, or one after the helper changed its shell's settings, does not count. The agent that launched the helper reads the verdict for each condition, and your answer shows one line with the counts: «Checked by the app: … hold, … do not hold, … not verified».

## Create a subagent type

1. Open «Settings» (⌘,) › «Agents & helpers».
2. Open «Advanced». Under «Kinds of helper» you see «Built in», «Project» and «Yours».
3. Press the «New type» icon, or the «Duplicate to edit» icon on an existing one. Hover an icon to read its name.
4. Fill in:
   - «Name»: letters, numbers, - or _. The agents call it by this name.
   - «When to use it»: what it does and when. The agents read this to choose it.
   - «Tools»: add tools, or keep «Every tool of whoever launches it».
   - «Effort»: «The model’s own», «Low», «Medium» or «High».
   - «Instructions»: what it does, how, and what it delivers.
5. Press «Save type». You see «Type “…” saved.»

To remove one of your types, press «Remove (it goes to the Trash)».

## If something goes wrong

- **«The name takes letters, numbers, - or _ (64 at most).»**: change the name.
- **«You already have a type called “…”.»**: pick another name.
