Cursor 2.4: Subagents

[what.gift]

Which models are setup to use it?

Do I have to prompt to say to use the “subagent” like your example video or it will automatically and intelligently do it for me?

I’m not sure when I should be saying to use it or not - some lack of direction and best practices for it?

[Colin]

Hey @what.gift!

You don’t have to explicitly tell Cursor to use the subagent. The model is aware of the subagents available and decides when to use them. You can, of course, nudge them in that direction by being more directive.

[liquefy]

In docs it says we can provide model id to the subagent, but where do we find model ids? docs and settings show only model names, not their ids

[Artemonim]

@liquefy

where do we find model ids?

https://cursor.com/dashboard?tab=usage, Model

[Colin]

The model picker within the subagent editor will also autopopulate the model for you!

[Artemonim]

It would be great to be able to open the “control room” to view all active agents. Maybe even as a separate window.

Example:

1. I launch the main agent with instructions to use Subagents.
   2a. Currently, if you don’t touch the PC, Subagent will appear as a small embed in the chat.
   2b. I’d like to explicitly open a new chat and automatically display the work of the Subagent(s). And so that the work coming from one main Agent feels like a whole group of agents, and not like another equivalent chat.

Honestly, I don’t really understand how this should look, but I hope I’ve conveyed the general vibe

[aamaralb (Alex)]

hey @Colin first of all thanks for this ! i see you are in the frontline and answering every question ! I have run in to something and i’m not sure if its intended behavior

Subagent Delegation Not Working from Parent Directory

Environment

• Cursor Version: 2.4.21 (arm64)
• OS: macOS (Darwin 25.2.0)
• Subagent Location: ~/MyCompany/project/.cursor/agents/
• Subagents Defined: developer.md, reviewer.md

Issue Description

I’ve configured two subagents in my project at ~/MyCompany/project/.cursor/agents/ (a developer agent and a reviewer agent). I’m experiencing inconsistent behavior with
subagent delegation depending on where I launch Cursor IDE:
Works: When I open Cursor IDE directly in ~/MyCompany/project/ and ask the agent to delegate to subagents, delegation works correctly.
Doesn’t Work: When I open Cursor IDE in the parent directory ~/MyCompany/ and ask the agent to delegate to subagents, I get generic responses like “I don’t have the
tools to delegate” instead of proper delegation.
Also Doesn’t Work: Cursor CLI doesn’t seem to delegate to subagents even when launched from within ~/MyCompany/project/.

Steps to Reproduce

1. Create subagent definitions in ~/MyCompany/project/.cursor/agents/
2. Open Cursor IDE in ~/MyCompany/ (parent directory)
3. Ask the agent to delegate a task to one of the subagents
4. Observe: Agent responds with “I don’t have the tools to delegate” or similar generic message

Compare with:

1. Open Cursor IDE directly in ~/MyCompany/project/
2. Ask the agent to delegate the same task
3. Observe: Delegation works as expected

Expected Behavior

Subagent delegation should work regardless of which parent directory the IDE is opened from, as long as the workspace contains the .cursor/agents/ directory. The agent
should be able to discover and delegate to subagents defined in any workspace subdirectory.

Questions

1. Is subagent discovery limited to the exact workspace root where .cursor/agents/ exists?
2. Should subagents be discoverable from parent directories, or is this the intended behavior?
3. Is there a way to make subagents available across a multi-project workspace hierarchy?
4. Why doesn’t Cursor CLI support subagent delegation even from the correct directory?

Additional Context

My project structure:

 1 │~/MyCompany/
 2 │├── project/
 3 ││   └── .cursor/
 4 ││       └── agents/
 5 ││           ├── developer.md
 6 ││           └── reviewer.md

This behavior makes it challenging to work with monorepo-style setups where you might want to open the IDE at a higher level but still leverage project-specific subagents.

Also, with the IDE open in parent directory of project/ if i explicitly ask the agent to work from MyCompany/project it then proceeds to invoke a general-purpose subagent over the defined developer agent

[Colin]

Hey @aamaralb

Let me see if I can address all your questions:

• Subagents, today, are discovered only at the workspace root - specifically <workspaceRoot>/.cursor/agents/
• There is no recursive search into subdirectories
• You can add agents to your home directory, ~/.cursor/agents/, which will be available for all projects.
• Subagents should be supported in the Cursor CLI as of last Friday’s release. Make sure you’re running 2026.01.23-916f423.

[Eddy]

AFter extensive testing:

• It worked on Nightly version of 2.4 before, up to 8 in parallel (better with Claude mdoels, gemni never worked, Codex was hit or miss, Composer doesnt even understand the query and just does its own thing)
• as of version 2.5 they no longer work at all, Irregardless of the model, it can find the Agents anywhere (even outside of workspace, found the claude ones and could run them in 2.4) knows how to call it via the internal “Task” tool, but the tool is not available and it cant figure out how to make it available (neither can i sadly…)
• Image generation doesnt work at all on 2.5 either, the models doesnt know of its existance (Opus 4.5 thinking had the idea to search online and found the docs but cant find the exposed tool - which one is it?)
• Claude skills AND command import is flawless and execution too, running multiple slash commands and skills in the same query is a FREAKING productivity boost, claude code terminal just CANT only one at a time (workaround is to explcitly tell it to “invoke SKill X with argument Y, works perfect then)
• funny enough Cursor models dont give a ■■■■ if skills or commands have any YAML formatting or not, it reads them and executes flawlessly regardless of model or mode (FREAKING dope Cursor team <3)

Platform windows 11, only cursor IDE didnt test in CLI

Hope to get the subagents working again <3

[Colin]

Hey @Eddy

If you’re facing bugs in 2.5 (nightly), definitely raise new threads over in Bug Reports!

[Eddy]

Thought you guys auto track all the crap we users do in the Nightly and Early access mode ^^

so didnt want to spam with every little thing when you guys are working on many features anyways, but will share that there if you want (or if it makes sense)

[Colin]

If you post something we’ve heard about, we’ll tell you.

And privacy mode is privacy mode whether or not you’re on Nightly or Stable, so we don’t always know everything!

[Moumouls (Antoine Cormouls)]

Can subagents call other subagents?

[Eddy]

never worked for me ^^ i tried xD, but they called CLI where you can call gemini-cli or qwen cli, so inception mode enabled ^^

[Colin]

Hey @Moumouls

@Moumouls

Can subagents call other subagents?

Moved your post in here because I think it’s a great question – no, subagents cannot call other subagents. This is done on purpose to prevent recursive subagent spawning.

[Eddy]

do you have checks to prevent calling COdex-CLI and Gemini-CLI=?

[Colin]

Nope!

[Eddy]

then nothing speaks against us using it if we wish? cause its not Token usage of Cursor right? (not sure it would be worth it though cause it will block the exection of the whole chatm so guess better start them in a regular terminal ^^)

[david_g]

Is it possible to assign a MCP only to subagents. That would lower the token used by with a more efficient context. For example, I could have an agent specialised to create github issues, PR and so on. The MCP would not be loaded in the main Agent (github is 40 tools). Specialised subagents would load the MCP but would require small context.

[Artemonim]

@Eddy

then nothing speaks against us using it if we wish?

I thought about using them as subagents by calling through the terminal, but I’m too lazy to test it yet.

I already know that Claude Code can be called with a command that will only accept the request, do the work, and give a final answer without interactive questions.