Sub-Agents

Sub-agents let the AI split work up. When a task is big or naturally parallel, the model can spawn one or more child Chat Nodes on your canvas — each sub-agent runs its own conversation independently and reports its final answer back to the parent, which then continues with the results.

You don't need to enable anything: just ask. Prompts like "run three sub-agents to research these topics in parallel" or "delegate this comparison to sub-agents" trigger the behavior on models that support tool calling.

---

How it works

1. The parent AI calls its run-subagent tool with a short label and a self-contained task for each delegation — multiple calls in the same turn run in parallel

2. A new chat node appears on your canvas, connected to the parent, and starts working immediately

3. The sub-agent inherits the parent conversation as connected-node context, plus its task instructions

4. When it finishes, its final answer flows back into the parent's tool result and the parent resumes

The AI can also read your canvas structure through a canvas-context tool, so it can see what nodes already exist when deciding what to spawn. See Core Concepts for how connected-node context works.

Depth is capped at 1 level: a sub-agent is not offered the run-subagent tool itself, so sub-agents cannot spawn their own sub-agents and delegation can't recurse indefinitely.

---

Running sub-agents on different models

Each sub-agent can run on a different model than the parent chat. Just name the models in your prompt and the AI assigns one per sub-agent when it spawns them:

"Debate whether we should rewrite the backend in Rust. Run three sub-agents to argue it — one on @grok-4, one on @gpt-5.5, one on @opus — then synthesize a verdict."

• Loose references work — "@opus", a bare model id, an exact provider-and-model pair, or the model's display name all resolve to the closest match.
• Only your enabled models can be picked — the AI is offered the models you've turned on in Models, plus the built-in Basic / Pro / Advanced tiers. If it names a model that isn't available, that sub-agent fails immediately with the list of valid options, and the parent can correct itself or ask you.
• Say nothing about models and behavior is unchanged — every sub-agent inherits the parent chat's model.
• The child node shows its model — check the spawned node's model picker to confirm which model is answering.

---

What you see while it runs

Inside the parent's tool-call card, a live activity card tracks each sub-agent:

• A status line — Starting subagent…, then Subagent is thinking… or Subagent is responding…, prefixed with the child node's label
• Running tool: the tool the sub-agent is currently using, when it makes its own tool calls
• A streaming preview of the last stretch of the sub-agent's response text
• Jump to node — pans and zooms the canvas to the child node so you can watch it work
• Cancel — stops that sub-agent's run

---

After the run

The child node persists on your canvas. You can:

1. Open it and read the full conversation, including every tool call it made

2. Continue chatting in it directly — it's a normal chat node

3. Keep it connected as context, or delete it if it was throwaway work

Runs can also end as failed, cancelled, or timed out — the parent is told the outcome either way and explains what happened.

---

Use cases

• Parallel research: "Spawn a sub-agent for each of these three competitors and compare their pricing" — three nodes work simultaneously, the parent synthesizes.
• Multi-model debates: run the same question past several models — "have @grok-4, @gpt-5.5, and @opus each take a position" — and let the parent weigh their answers.
• Long side-quests: delegate a deep document analysis to a sub-agent while the parent conversation stays focused on the main thread.
• Auditable delegation: unlike hidden background steps, every sub-agent is a visible node — you can inspect exactly how it reached its answer.
• Iterating on a delegation: if a sub-agent's answer is close but not right, jump to its node and continue the conversation there instead of re-running everything.