Sub-agents
Let the AI split a big question into parallel tasks that each run in their own chat node and report back.
Big questions are often several smaller ones. Sub-agents let the AI hand each part to its own chat node, run them side by side, and bring the answers back to the chat that asked. Nothing is buried in one long reply: every sub-agent is a node you can read, check and keep talking to.
When the AI uses sub-agents
Sub-agents are part of Agent mode, the default. Chat and Image modes don't use tools, so they never start them. Orchestrator mode uses them too, for research it then turns into nodes.
Ask in plain words: "research these three companies in parallel". The AI can also split a task on its own when the parts are independent.
Tip: If you want nodes on the canvas as the result (say, one per topic), Agent mode proposes switching to Orchestrator mode instead. See Agent mode and tools.
What you see
- A chat node appears for each task, connected from the parent and stacked in a column to its right. Its first message is the task the AI wrote. Being connected, it also reads the parent's conversation, like any connected chat.
- In the parent, each sub-agent is a Subagent row (several are grouped under a heading like "Ran 3 subagents"). Open a row to follow it live: its status, the tool it's using, the latest lines of its reply, Jump to node and Cancel.
- Each sub-agent reports its final answer back. Once all have, the parent carries on, usually by combining them.
Choosing models
Sub-agents use the parent chat's model unless you name others: "have three sub-agents argue this, one on Claude, one on GPT and one on Gemini." Loose names work. The AI can pick Auto, any Slashspace hosted model, or any chat model in your model picker. If a name matches nothing, that sub-agent fails before a node is created and the parent is told which models it can use.
Each sub-agent is billed like any chat on its model: hosted models (including Auto) draw from your hosted wallet; your own API keys, subscription providers and Ollama don't. Since each one reads the parent's conversation, a long parent thread makes every request bigger. See Plans and billing.
Limits
| Limit | Details |
|---|---|
| Depth | One level. Sub-agents can't start their own, even when you chat with them later. |
| Time | 10 minutes each, then it's stopped and reported as timed out. |
| How many | No fixed number. They all run at once. |
| Stay on the canvas | Switching canvases or quitting mid-run marks the run interrupted. Ask again if you still need it. |
If a sub-agent's tool needs approval, the card appears on its own node while its row says "Subagent is waiting to continue…". Approve it there before the time runs out.
Stopping and retrying
Cancel on a row stops that sub-agent, as does stopping the chat on its node. Cancelled runs don't wake the parent, so if you cancel them all, send the parent a message to carry on.
When a sub-agent fails or times out, the parent is told why and can retry it in the same node, on the same model or a different one. You can ask too: "retry the one that failed on another model."
Claude Code, Codex and Cursor
Subscription providers run tools inside their own agent, so Slashspace adds its sub-agent tool there and asks the agent to use it instead of built-in helpers that run out of sight. Each sub-agent is still a connected chat node with the same limits, and its answer goes back into the session. The parent shows a Subagent row without the live preview, so watch the child node. Stop on the parent cancels its sub-agents.
Continuing a sub-agent's chat
A sub-agent node stays on the canvas as a normal chat. Open it to see the whole run, tools included, then keep talking to it, branch from it, or delete it.
After your next message in the parent, earlier sub-agent answers shrink to short excerpts there. The full answer stays on the sub-agent's node, so connect that node to any chat that needs all of it.