> ## Documentation Index
> Fetch the complete documentation index at: https://docs.archetype.ing/llms.txt
> Use this file to discover all available pages before exploring further.

# Agents

> Child agents that split the work, agents you name, orchestration, and sessions that talk to each other.

## Child agents

The agent can hand parts of a request to child agents. Each starts fresh, works on its part, and gives back only its final answer. Several run at once. You see a line for each child under the step that started it, and an approval a child asks for says which child it is.

| Kind | What it does |
| - | - |
| Inspect | Reads only. |
| Edit | Works in a temporary Git worktree of your folder. When it ends, its changes are merged into your folder, not staged or committed. If they conflict, nothing is written and the patch is kept. Needs a Git repository. |
| Background | The agent goes on while the child works; the result comes back to the conversation when it is done. |

A child never gets more than its parent may do: your [permission mode](https://docs.archetype.ing/en/use/approvals.md) and approval rules apply to its calls too. A child that has a question asks its parent first; if the parent does not answer within 10 minutes, the question comes to you and shows under **Needs you**.

Child agents are on for Archetype's own models, Claude Code and Codex. Turn them off with `"children": false` in your user config.

**Tokenmaxxing** puts speed first: the agent splits the work into independent parts and runs them all as children at once. Turn it on from the status bar or with `/tokenmaxxing`.

## Named agents

Two agents come built in: `general`, and `explore`, which only reads. Add your own as Markdown files; the agent picks one by its name and description.

```markdown theme={null}
---
name: reviewer
description: Reads the changed files and finds bugs and missing tests
mode: inspect        # inspect or edit; without it, the task decides
tools: [read_file, search_text]   # without it, every tool of the mode
model: gpt-6-luna    # a model of the parent's provider; without it, the parent's
max-turns: 30        # 1 to 200; without it, no limit
---

Read the change closely. For each problem, give the path, the line and why.
```

| Where | For |
| - | - |
| `~/.config/archetype/agents/*.md` | You, in every folder |
| `.archetype/agents/*.md` | This project |

Yours replaces a built-in agent of the same name. A project's agent only adds new names: one named like a built-in or yours is skipped with a warning, so a cloned repository cannot change what `explore` does.

## Orchestration

For work bigger than one conversation, one session leads others.

**1. Start it**

Type `/orchestration` and what you want, or choose **Orchestrate** in a folder's new-session menu.

**2. Agree to the plan**

The lead writes a plan: the steps, what done means for each, and which engine does it. Nothing starts until you answer "Start". Ask for changes in words and the lead rewrites it.

**3. The steps run**

Each step runs on its own branch, up to 3 at a time. When a step is done, another engine checks it if one is signed in. A step that fails its check twice waits for you.

**4. Take the result**

By default it ends with each step's pull request ready and its checks passing; the lead proposes the order to merge them.

Questions from a step go to the lead first; what only you can decide comes to the lead's conversation. Settings are under `orchestration` in your config; `"orchestration": false` turns it off.

## Sessions that talk to each other

Conversations on the same computer can message each other (over the [A2A](https://a2a-protocol.org/) protocol). Ask the agent to check with another session, or to start work in another folder.

* A message to a session that is working joins its current request; otherwise it starts a new one there.
* The receiving session works with its own permissions, and asks its own approvals.
* Sending and stopping another session asks first in `ask` mode, and is refused in `read-only`.
* The agent opens a new session only when you ask it to; otherwise it splits work into child agents.
* After 8 messages back and forth with no person, no more are sent.

Turn it off with `"a2a": false`; in a project's config it turns it off for that folder only.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.