# Agents that start sessions

> Let an agent ask you questions, notify you, and split big work into sessions of its own.

Source: https://sesh.codes/docs/agents-that-start-sessions/

Sesh gives the agents it runs a few tools of their own. With them an agent can ask you a question wherever you are, send you a notification, show you a chart, and start other sessions and follow them, so one conversation can run several agents at once.

## What agents can do

| Ask the agent to… | What happens |
|---|---|
| Ask you before it decides | The question shows in the chat, in **Needs You**, on your phone and watch, with its options as buttons. |
| "Ping me when the tests pass" | It sends you a notification in its own words. See [Notifications](https://sesh.codes/docs/notifications.md#notes-from-the-agent). |
| Show you something | It makes a chart, table or page in the chat. See [Artifacts](https://sesh.codes/docs/artifacts.md). |
| Rename the session | It renames the session, and its branch while it isn't pushed. |
| Split the work up | It starts sessions of its own. See below. |

You don't need to name the tools. Ask in plain words, and the agent picks them up.

## Turn them on or off

Both are on by default, per machine.

1. Open **Settings › Machines**, pick the machine and click **Details…**.
2. **Sesh Tools for Agents** turns all of Sesh's tools on or off.
3. **Let Agents Manage Sessions** lets agents start, message and stop other sessions. It needs Sesh Tools for Agents.

Turning them off takes effect for every agent on that machine straight away.

## Split work into sessions

Ask for something that splits naturally: "Do the API change in ledger-api and the UI in tidepool-web, in separate sessions, and tell me when both are done." The agent can:

- start a session in any project on the machine, with a task, and with another agent if you ask for one
- send a session a message, or stop its turn
- wait for its sessions to finish and read what they said

Each session it starts gets its own branch and worktree in a git project, like one you started. When one finishes, fails or needs you, its parent gets a short report and carries on from there.

> **Tip.** Agents use their own subagents for quick research. They start a session when the work should land on its own branch, run for a while, or use another project or agent.

## Follow along

- Sessions an agent started sit under it in the sidebar, indented. The parent's row counts them ("2/3 working"); click the count to fold them away.
- In the parent's chat, each step that touched another session names it.
- **Go to Parent** is in a child session's menu.
- A message from another session says which one sent it, so you can tell it from yours.

## Approvals stay with you

An agent can't answer another session's approvals or questions. They wait in **Needs You** for you, as usual.

Claude Code asks before an agent starts, messages or stops a session, unless the session is **Autonomous**. Looking at sessions doesn't ask.

## Good to know

- An agent can have up to 8 of its sessions working at once.
- Sessions an agent started can't start sessions of their own.
- An autonomous agent's sessions are autonomous too, each in its own worktree.
- You don't get a "finished" notification for a session that reports back to its parent; the parent hears instead. Its approvals and errors still notify you.
- Archiving a session that started others asks whether to archive them too.
- Agents need to support MCP tools to use these. Claude Code and Codex do.
