Do sub-agents work for you?

[mystilleef]

I know sub-agents are still experimental but I've never been able to trigger them, even though they seem to be properly refreshed and listed in /agents list.

Even when I ask for direct invocation of a sub-agent via a custom command the main agent just simply reads the agent file in the agents folder, and executes the instructions instead of delegating to the sub-agent.

I have the experimental flag enabled in settings. I'm running the nightly build. All the agents in the agents folder are registered and listed. I've even updated my SYSTEM.md and GEMINI.md to make them sub-agent aware.

Am I missing something? Anyone else got sub-agents working?

[gundermanc]

Hello, this feature is still under active development. For best results, please try the following:

• In settings, enable Preview features.
• Change your model to Gemini 3 pro preview or Gemini 3 flash.
• Enable the subagents experimental feature.
• Restart the CLI.

[mystilleef]

Thanks for your response. Still no luck. I already had preview features enabled, and my model set to Auto (Gemini 3). I've also manually set the model to the Gemini 3 Flash. I think I'll revisit once development has settled.

[hejnal]

Try adding the agent to the allowed tools in your settings.json file (with the same name). That worked for me.

eg:

"tools": {
    "allowed": [
      "security-auditor"
    ]
  }

[mystilleef]

Try adding the agent to the allowed tools in your settings.json file (with the same name). That worked for me.

eg:

"tools": {
    "allowed": [
      "security-auditor"
    ]
  }

Try adding the agent to the allowed tools in your settings.json file (with the same name). That worked for me.

eg:

"tools": {
    "allowed": [
      "security-auditor"
    ]
  }

Wouldn't that disable other tools?

[hejnal]

I don't think it would disable anything else (other tools might still ask for the user permission, at least when running gemini cli in the interactive mode). Probably worth trying it anyway.

I think the more modern way to manage the permissions is to use the policies
https://geminicli.com/docs/core/policy-engine/

[aniruddhaadak80]

From my point of view, the fact that the main agent reads the agent file instead of delegating suggests the feature boundary is still not stable enough for users to rely on prompt-only invocation patterns. The maintainer guidance in the thread is probably the practical answer for now, but the product would benefit from a clearer signal showing whether a sub-agent was actually invoked versus merely referenced. Without that visibility, it is hard for users to tell configuration problems apart from product limitations.