Agents working together
An agent can open other agents, give them tasks and wake up when they are done. It works on the canvas, within the limits you set.
What an agent can do
Agents in Ruimte use the ruimte-context command; ruimte-context help in a Ruimte terminal lists what it offers. With it, an agent can:
- read the nodes with a line into it, and the views and nodes you referenced with
# - add notes, terminals, chats, browsers and groups to its canvas, and draw lines between them
- open other agents and give them tasks
- send a message to another terminal or chat of the project
- drive a browser or device it has a line to
- publish results to Output
- add launches and routines, which wait for your approval
Delegate a task
Ask an agent to split up the work. It opens an agent per part beside itself, with a line into each that stays dashed while the task is open.
The parent wakes with each result. Tasks given in one go wake it once, after all of them finish.
| Rule | Detail |
|---|---|
| Permission mode | A child never gets a wider mode than its parent. |
| Depth | Agents nest at most two levels below the one you started. |
| Done | A task is done only when the child finished everything, background work included. |
| Usage limit | A child that stops on a usage limit pauses its task. The parent waits until the child finishes after the reset. |
| Questions | A child's questions and approvals come to you. Its parent may answer a question, never an approval. |
| Stopping | Stopping or deleting an agent ends the agents it opened. Their nodes and work stay on the canvas. |
What wakes a chat
An AI chat that is not working wakes up for:
- a task it gave that finished
- a message from another agent
- a pull request you asked it to watch, when checks finish or a review comes in; see Pull requests
Messages between agents stop waking each other after a few steps. A terminal is never woken. A terminal agent hears a message at the start of its next turn, or sees it printed on its screen.
Worktrees
An agent can open a child in a worktree: a separate checkout on a new branch. The child's node shows its branch. When the child is done, the parent can merge the worktree into a branch without uncommitted changes. A conflict goes back to the parent to resolve.
The worktree and its branch stay until you remove them. See Worktrees.
Let a chat arrange a canvas
A chat that is a view of its own can lay out a canvas you name. Ask it to set up a canvas; it adds nodes, draws lines and uses the browsers and devices there.
Set what agents may do
Settings, Agents, What agents may do, per machine:
| Setting | When on |
|---|---|
| Agents may delete any view or node | Agents can delete any view or node. When off, they delete only what they created. |
| Agents may reach all open projects | Agents can read your other open projects, message their terminals and chats, and start agents there. |
| Continue after a usage limit | See Usage limits. |
| Visual replies | See Visual replies. |