@Colin@deanrie Until the last 1 hour or so, I have been able to ask Claude Opus 4.6 High Thinking to launch sub-agents (I have many). Now I am getting an error from the agent indicating it does not have the “Task” tool available, which is required for launching sub-agents. It seems like the internal prompt is just not getting the correct tool list. Since Claude Opus is one of the few models that can actually use sub-agents, it is absolutely critical that this is working.
Steps to Reproduce
open a new tab
choose the claude-4.6-opus-high-thinking (or even claude-4.6-sonnet-medium) model
ask it to launch a sub-agent
ask it why it cannot launch the subagent
Expected Behavior
it should be able to launch the sub-agents, since it is one of the few models that can actually do this typically. it seems like a missing tool declaration for the “Task” tool
I’ve been looking into this on our end. The ability to launch sub-agents is controlled server-side, and based on your version (2.6.20), it should be available. I wasn’t able to pinpoint a specific change on our side that would have caused it to disappear.
To help narrow this down, a couple of questions:
Is this still happening? A similar report from another user resolved itself after about a day. If you restart Cursor (fully quit and reopen) and start a fresh conversation, does the Task tool appear?
If it’s still happening, could you grab a Request ID from a conversation where the Task tool is missing? You can find it by clicking the “…” menu on any message in the Chat sidebar and selecting Copy Request ID. This lets us trace the exact request on our end and see why the tool wasn’t provisioned. (Request IDs don’t contain any code or prompt data, just a lookup key.)
Is this affecting all models or specifically Claude Opus 4.6 and Claude Sonnet 4.6?
Yes this is still happening. I have restarted Cursor, updated Cursor to the latest version multiple times now, started new conversations, and tried it in a completely fresh and empty project. It still happens. This is how the thread goes (as well as every thread with any model i have available):
Me:
Do you have the ability to launch my subagent /product-owner with a simple request like “write a plan to add a header bar to the app”
Agent:
Unfortunately, no — I don’t have the ability to launch subagents in my current session. While the system context references delegating to a product-owner subagent via a Task tool, that tool isn’t actually available in my current toolset. The tools I have access to are file operations, shell commands, search, web fetch, MCP tools, and a few others — but not subagent delegation.
This is likely a Cursor limitation in the current mode. A few options:
I can do the work directly — e.g., if you want a plan for “add a header bar to the app,” I can switch to Plan mode and draft that for you myself.
You could check if subagent support requires a specific Cursor configuration — the Task tool may need to be enabled or may only be available in certain Cursor versions/modes.
Would you like me to go ahead and help with the header bar plan (or whatever task you had in mind) directly?
I can grab this, but it probably wont be accessible to you. My company, at the global level, sets all our accounts to use Privacy Legacy, and that is unchangable for me. Here is the id for the thread above: 67d418ca-b51b-4215-8ede-54205776b25b
I have tested over 10 models (the top capable models) and none of them allow it, whether this is GPT, Claude, or Gemini.
thanks for the detailed info, and sorry for the delayed follow-up.
I’ve traced the root cause. On your team’s plan configuration, subagents require a specific internal model (Composer) to run. If your team admin has restricted that model via the model blocklist, the Task tool won’t be available, which is why subagents can’t launch regardless of which model you select in the chat.
Two paths forward:
Ask your Cursor team admin to check the model configuration and ensure Composer isn’t blocked. That should immediately restore subagent functionality.
Try Max Mode – when Max Mode is enabled, subagents use the parent model instead and bypass the Composer requirement. Note that Max Mode sessions count against your request budget.