Ability to explicitly choose an subagent

[quocvu]

What would you like to be added?

Thank you for adding subagents @tanzhenxin. It's a really a huge improvement over Gemini CLI.
I noticed the AI doesn't often delegate to subagents. It could be my description of the subagents being not as comprehensive. Is it possible to have more guidance on how to increase our chances of triggering subagents? Or I may suggest adding a way for user to explicitly force the AI to delegate to a chosen agent. Maybe in a form of a command.

Why is this needed?

A more predictable way to assign tasks to subagents.

1. we don't want the AI to pick the wrong agent wasting time and tokens
2. too often, the AI doesn't delegate and render this feature inaccessible

Additional context

No response

[tanzhenxin]

You can try using a prompt in this format:
“Use [Agent Name] to do [Task Description].”

This helps the main agent recognize and select the mentioned subagent.

---

Keep in mind that we can only prompt the model — we can’t directly control its behavior.
Ultimately, the model decides which tool or subagent to use, and we’re continually working to improve its ability to choose and use tools effectively.

[pomelo-nwu]

@tanzhenxin Could we add a command like /agents use {subagent} to enable users to invoke subagents as conveniently as they do with @memory-file?

[tanzhenxin]

@tanzhenxin Could we add a command like /agents use {subagent} to enable users to invoke subagents as conveniently as they do with @memory-file?

The subagent feature works more like an additional tool — the model decides when and which subagent to use during task runtime.
In contrast, the memory file functions as a task setup feature — its content is concatenated into the prompt before execution begins.

We should avoid forcing any specific task runtime behavior on the model, as that’s not the right path forward.

[pomelo-nwu]

@tanzhenxin I completely understand your point. I used @memory-file as an example because the @ symbol naturally surfaces a list of available options, making it effortless for users to select one.

Similarly, /agents use could behave the same way — displaying a dynamic list of all globally and locally defined agents, helping users quickly pick the right one when they know what they want to do but can’t recall the exact agent name.

You can think of this as a convenient shortcut: it serves the same practical purpose as typing out a full prompt like “Use [Agent Name] to do [Task Description].” — just faster, more intuitive, and less error-prone.

[tanzhenxin]

@tanzhenxin I completely understand your point. I used @memory-file as an example because the @ symbol naturally surfaces a list of available options, making it effortless for users to select one.

Similarly, /agents use could behave the same way — displaying a dynamic list of all globally and locally defined agents, helping users quickly pick the right one when they know what they want to do but can’t recall the exact agent name.

You can think of this as a convenient shortcut: it serves the same practical purpose as typing out a full prompt like “Use [Agent Name] to do [Task Description].” — just faster, more intuitive, and less error-prone.

We may need to reconsider the nuance here.
If we provide a slash command like @sub-agent, users might expect the sub-agent to be invoked deterministically — as if it’s guaranteed to run — but in reality, we don’t have that level of control.

In contrast, the prompt style “Use [Agent Name] to do [Task Description].” works more like a suggestion rather than a direct command, which better reflects how the system actually operates.

[pomelo-nwu]

@tanzhenxin 「but in reality, we don’t have that level of control.」Got it — I have no further questions on this. Thanks for your help!

[quocvu]

@tanzhenxin thank you for the clarification. I understand it is up to AI to pick the agent as said in the docs.
But given it rarely does, it diminishes the impact of this feature. It is the case with some tool as well. I rarely see it invoke sequential thinking and context7.
May be you could suggest to the core team to make sub agents more deterministic.

On another note, I stopped using sub agents because the screen flickers when the agent is in action.

[wanyukang]

@tanzhenxin thank you for the clarification. I understand it is up to AI to pick the agent as said in the docs. But given it rarely does, it diminishes the impact of this feature. It is the case with some tool as well. I rarely see it invoke sequential thinking and context7. May be you could suggest to the core team to make sub agents more deterministic.

On another note, I stopped using sub agents because the screen flickers when the agent is in action.

The actions that trigger MCP server are very rare, which seems to be caused not only by the model's poor ability to follow instructions. If this is not optimized, the MCP server will become like a decoration for qwen-code and have no practical value.

[geraldthewes]

I'm having the same issue and not quite sure what to do. I created an agent to encapsulate the knowledge of a custom build tool being used by a project, and most time when i prompt it, I get this response from Qwen (granted qwen 30B 4 bit)

"However, I notice that I'm not able to directly execute the jobforge manager agent in this environment. Let me try another approach to help you with your request."

And then gets lost trying something else I don't want to. It's really hard to get to use it. No sure if I should abandon the agent route and put the same information in QWEN.md, but sort of makes these agents not very useful.

[wanyukang]

After upgrading to version 0.4.0, I've noticed a significant increase in the chance of triggering MCP. I'm not sure whether it's due to improvements in the CLI capabilities or a silent update to the base model, but overall, the performance and user experience in the current version are much better than before. fantastic improvement! Great job, team!