really? the faq in docs says:
The built-in computerUse subagent cannot spawn additional subagents; the Task tool is explicitly filtered out. Custom subagents inherit all tools from the parent by default, including the Task tool, so they can spawn other subagents.
and if i understand correctly that means subagents can spawn other subagents?
I just tried asking Grok Code Fast to call a Grok Code Fast subagent which would call a Grok Code Fast sub-subagent, but the subagent didn’t call the sub-subagent.
It would be great if “system” subagents could be disabled or at least visible in the settings. I like explorer, but I’d like more control.
hey @Colin Thanks,
Subagents, today, are discovered only at the workspace root - specifically
<workspaceRoot>/.cursor/agents
This only for the IDE right ? so that have the IDE detect any subagents definition, i need to open that folder as the root folder in the ide?
Subagents should be supported in the Cursor CLI as of last Friday’s release. Make sure you’re running
2026.01.23-916f423
I am running that version but i still cant detect my subagents def.
i open agent in ~/MyCompany/projecte.g:
[aamaralb@AAMARALB-M-D6P9 ~/MyCompany/project]$ agent
Cursor Agent v2026.01.23-916f423
~/MyCompany/project
(cwd is not a git repository, cursor rules and ignore files don't apply)
┌──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┐
│ invoke the developer agent with prompt "write me a simple number guessing game in python" │
└──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘
I don't have a tool available to invoke a "developer agent" directly. However, I can help you create a simple number guessing game in Python right now.
Let me create this for you.
@aamaralb What I think is going on is that you are on a request-based plan, and right now it requires having Max mode enabled to use subagents. That will change very soon (we’re rolling out some changes).
In the meantime, try turning /max-mode on.
Thanks i thought about that because i also had to turn it on in the IDE as per docs, but no success 
[aamaralb@AAMARALB-M-D6P9 ~/MyCompany/project]$ agent
Cursor Agent v2026.01.23-916f423
~/MyCompany/project
(cwd is not a git repository, cursor rules and ignore files don't apply)
┌─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┐
│ invoke the developer agent with prompt "write me a simple number guessing game in python" │
└─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘
I don't have a tool to invoke a "developer agent" directly. However, I can help you create a simple number guessing game in Python myself.
Let me create that for you.
⬢ Listed . 3 files, 2 directories
I see there's already a number_guessing_game.py file and what appears to be agent configurations. Let me check the existing file and the developer agent definition.
⬡ Reading 2 files
Reading number_guessing_game.py
Reading .cursor/agents/developer.md
┌─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┐
│ → Add a follow-up │
└─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘
Claude 4.5 Opus · MAX · 16.5%
/ commands · @ files · ! shell
I wrote up a showcase to demonstrate multi-agent and published the github repository.
@Colin I found in documentation that there should be 3 built-in subagents (Explore, Bash and Browser - Subagents | Cursor Docs), but when I tried to call them like, only Task subagent (one or multiple in parallel) are executed, could you please explain which pre-built subagents exist now and how to call them properly?
Hey @labdmitriy!
Great question. The agent is responsible for deciding when to use these. It’s not possible to manually invoke them.
You don’t need to configure these subagents. Agent uses them automatically when appropriate.
@Colin Thank you for your response!
Could you please provide example to see that they triggered, and how to check it?
I see that all explore, bash and browser tasks are using the session’s context and it is not triggered at all, and I am asking to use separate subagent for such tasks to keep my parent context clean. So it would be great if you can clarify this behavior for built-in subagents and be sure that our custom agents will not conflict with non-visible built-in subagents invocation.
@Able1991 I have same experience. I even attached subagentStart/End hooks and they don’t get triggered.
Did you figure it out or post on the bug reports?
which model are you using ?
Hey @labdmitriy!
All subagents are visible when they’re used. What model are you using?
3 posts were split to a new topic: Why can’t I specify custom models for subagents?
No sure this has been answered before (search hasn’t shown me anything) but why aren’t subagents working in “auto” mode??
Hey @cyborg31
We hope to have that resolved soon.
How should sub-agents in Cursor be used when parallel edits might overlap, especially during large refactors? For example, if refactoring a parent module and its submodules and related tests, sub-agents will often need to update the parent’s signatures. How do we prevent conflicts if multiple sub-agents attempt to edit the same parent file simultaneously during such heavy-handed changes?
is there any reason why i don’t see the billing for subagents? like i’m using claude to fire off subagents, and i see it all under claude on my billing like it’s charing me for claude and nott the subagents which are way cheaper?
what i mean is i putposely set the subagent to use different models like grok code, or gpt mini, or gpt 5.2 etc, and when it does it, it shows up on my billing as claude still.
I know it’s not claude in the subagent because i can see its response and I’ve been working with the agents long enough to know their style of responses.