In AI Integrator Code, the coding agent leading a chat can now hand part of the work to a subagent that runs on a different coding program, model and effort, and that keeps working on Auto without waiting for you. The chat's plan stays pinned on one line under its header, and any model can stop to ask you a question on a card with choices to pick from.
What is a subagent in AI Integrator Code?
A subagent is a second agent that the chat's lead agent starts in the background to do a self-contained piece of the task, such as writing tests while the lead changes the code. It reports back to the chat when it is done. Each one gets a row in the chat's Subagents tab, where you can read its transcript, see its plan, and message or stop it.
A subagent can run as one of your Agents, bringing that Agent's instructions, Skills and connectors. It can also be a one-off, whose role and instructions the lead writes for that task alone.
Can a subagent use a different model or coding program?
Yes. The lead can start a subagent on any coding program installed on your computer, such as Codex, Claude Code, OpenCode, Grok, Cursor or Antigravity, with its own model and effort. A Claude Code chat can ask a Codex subagent for a second opinion, or send a large, simple job to a faster model. When the lead names none, an Agent uses the program and model set on it, and a one-off uses the chat's.
It can also run on AI Integrator itself, which bills the model to your plan, as Code on your AI Integrator plan explains. If the program lists its models and the named one is not among them, the subagent stops, and the error lists the models it offers.
Subagents are set on the Agents page in Settings, under Chat & Agents, and in the Subagents menu of the Code composer. Integrator subagents is on by default when two or more coding programs are connected, and Max at once sets how many subagents can run together, from 1 to 20 (6 by default). Native subagents, on by default in the same place, lets the coding program also start its own subagents the way it does outside the app.
What happens when a subagent needs permission?
A subagent runs on Auto unless the lead starts it read-only, or the chat itself is read-only. On Auto it edits files in the project. A request that reaches past the project, such as a command the coding program asks about, the network, or a file outside the folder, is answered in order:
- The decision model, the same one Auto uses in a chat, decides when it can.
- If it cannot, the lead agent decides, and asks you first when it is unsure.
- If nobody answers within 10 minutes, the request is declined.
- Subagent asks. Past the project
- Decision model. Decides when it can
- Lead agent. Asks you if unsure
- Allowed or declined. Noted in its transcript
- Declined. No answer in 10 minutes
- Subagent asks to Decision model
- Decision model to Allowed or declined: Decided
- Decision model to Lead agent: Unsure
- Lead agent to Allowed or declined: Answered
- Lead agent to Declined: 10 minutes
A subagent never waits on a card that nobody is watching, and every decision appears in its transcript. It is never allowed more than the chat that started it. Antigravity cannot wait for the lead, so an Antigravity subagent refuses anything past the project instead. Plugin permissions in Chat and Code explains how Auto decides.
Where is the agent's plan?
On one line under the chat's header. It shows the step being worked on, how many steps are done ("2 of 5") and a thin bar that fills as steps finish. The chat scrolls under it, so the plan stays in view.
Click the line to see every step: a check for done, a dot for the one in progress, and an empty circle for the rest. A step with a small arrow has a sentence or two on what it covers. Escape or a click outside closes the list.
The plan looks and works the same in Codex, Claude Code, OpenCode, Grok, Cursor, Antigravity and the AI Integrator runtime, on the desktop app, the web and iPhone. The agent posts it with its first step, updates it as it works, and sends a new one when you change direction, so the line never shows a plan it has dropped.
How does a model ask me a question?
Any model in Chat, an Agent's Home or a Code chat can ask up to four questions on a card. Each question has two to four choices, with the recommended one marked, and you can pick one, check several, or type your own answer.
- A single question with two short choices shows as two buttons.
- A longer card steps through its questions with Next, Back and Send answers.
- Skip sends the recommended choices.
- While a card waits, whatever you type in the composer answers it.
In Code, a question holds the agent for up to 10 minutes (45 seconds in Cursor). After that the agent ends its turn, the card stays open, and your answer reaches it as your next message. A scheduled run, and any subagent, takes the recommended choices instead of waiting. Cursor still shows its own question card.
Subagents on their own coding program and model arrived in AI Integrator 0.3.5, the pinned plan and question cards in 0.3.9, and subagents on Auto in 0.3.10. The 0.3.5, 0.3.9 and 0.3.10 patch notes list the rest.


