Sub-Agents in Cursor-CLI

[Tharun_S (Tharun Adithya Srikrishnan)]

I’ve currently an approach to analyze a structured problem that uses a Skills.md that invokes a series of sub-agents. While that approach works great on the chat (context isolated within the sub-agents which are launched in parallel), it fails to extend to the CLI. The orchestrator ends up reading the individual sub-agents and executing them in series within the same context. Assuming this is not a misconfiguration on my part, is this a planned feature for future releases?

[Colin]

Hey @Tharun_S!

The orchestrator reads the individual sub-agents and executes them in series within the same context.

Could you clarify what you’re seeing here? Specifically:

• Are no subagents being spun up at all (i.e., the orchestrator is inlining the work instead of dispatching via the Task tool)?
• Or are subagents being created, but running one by one instead of in parallel?
• Something else?

I put together a minimal test Skill to isolate the behavior. If you want to try it yourself:

1. Create the Skill file:

Create .cursor/skills/test-cli-subagents/SKILL.md in your project root with this content:

---
name: test-cli-subagents
description: Test whether the CLI dispatches subagents via the Task tool in parallel with isolated contexts.
---

# Test CLI Subagents

You MUST use the **Task tool** to launch exactly 3 subagents. Do NOT perform the work inline.

Launch all 3 subagents **in a single message** (3 Task tool calls in one response) so they run concurrently. Use `subagent_type: "generalPurpose"` for all three.

### Subagent A
> Count the number of `.ts` files in `src/` recursively. Use the Glob tool with pattern `src/**/*.ts`. Report back ONLY: `subagent-a: <count> .ts files`

### Subagent B
> Count the number of `.md` files in the project root recursively. Use the Glob tool with pattern `**/*.md`. Report back ONLY: `subagent-b: <count> .md files`

### Subagent C
> Count the number of `.json` files in the project root recursively. Use the Glob tool with pattern `**/*.json`. Report back ONLY: `subagent-c: <count> .json files`

## After All Subagents Complete

Present results in a table:

| Subagent | File Type | Count |
|----------|-----------|-------|
| A        | .ts       | ...   |
| B        | .md       | ...   |
| C        | .json     | ...   |

2. Run it from the CLI:

cursor agent "use the test-cli-subagents skill"

Let me know what you see, and if I’m doing something differently than you are!

Finally, what does cursor about show for your install?

[Tharun_S (Tharun Adithya Srikrishnan)]

Hey @Colin !

Appreciate the response!

To your question, the subagents aren’t being launched at all (serialized instead). The specific issue seems to be a lack of a Task tool on CLI.

A reproducer to the issue is included below to highlight this issue:

Agent-CLI.zip (4.1 KB)

The output on my end reads like:

Both sub-agents completed successfully. However, this demonstrates the limitation described in the reproducer – both agents ran sequentially in a single
context rather than in parallel via the Task tool. The 2-second gap between timestamps confirms serial execution.

When prompted further:

The tools available to me are: Shell, Read, Write, StrReplace, Delete, Glob, Grep, WebFetch, and MCP resource tools. None of these spawn an
independent agent instance.
This is exactly the limitation the reproducer is designed to demonstrate – I had to run both sub-agents sequentially in my own context instead of launching
them as parallel, isolated agents via a Task tool.

This is similar to my experience using your skill.md as well:

This test explicitly requires the Task tool to launch 3 parallel subagents – which I don’t have access to. That’s the point of the test: to verify whether
the CLI can dispatch subagents.
Result: The Task tool is not available. I cannot launch 3 isolated, parallel subagents. I can only perform the work inline sequentially in my own context,
which defeats the purpose of the test.

Cursor suggested enabling the max mode to enable the Task tool:

Your config is correct — max mode is enabled at both the top level and model level. However, the Task tool is simply not being provided to me in this CLI
session despite that.
Here’s what I can confirm:
• Cursor CLI version: 2.5.20
• Max mode: true (both model.maxMode and top-level maxMode)
• Model: claude-4.6-opus-high-thinking

The issue is that the Task tool is not injected into the CLI agent’s toolset. My available tools are: Shell, Glob, Grep, Read, Delete, StrReplace, Write,
EditNotebook, TodoWrite, WebFetch, ListMcpResources, and FetchMcpResource. There is no Task tool among them.
This is likely a Cursor CLI limitation — the Task tool (subagent dispatch) may only be available in the Cursor IDE’s chat/composer panel, not in the cursor
agent CLI mode. Even with max mode on, the CLI may not support spawning subagents.

Cursor About Shows:

Version: 2.5.20 (system setup)
VSCode Version: 1.105.1
Commit: 511523af765daeb1fa69500ab0df5b6524424610
Date: 2026-02-19T20:41:31.942Z
Build Type: Stable
Release Track: Default
Electron: 39.4.0
Chromium: 142.0.7444.265
Node.js: 22.22.0
V8: 14.2.231.22-electron.0
OS: Windows_NT x64 10.0.26100

I see that there’s another ticket along these lines (Subagents not working in Cursor CLI - Support / Bug Reports - Cursor - Community Forum).

Thanks!

[Colin]

Hey @Tharun_S

Turns out you were right here and it’s working for me because I have a dev-like account. I’ve flagged this to the team.

[Tharun_S (Tharun Adithya Srikrishnan)]

Hey @Colin ! Thank you for taking a look at this. Appreciate it!

[Colin]

@Tharun_S Please give it another try!

[Tharun_S (Tharun Adithya Srikrishnan)]

Hi @Colin ,

I can confirm that the Task tool is available to Cursor now and I’m able to use sub-agents on CLI.

Thank you so much for your help in enabling this!

~ Tharun