Subagent Reasoning level

Is there any way for an agent to pass reasoning level to invoke_subagent or set it via ~/.gemini/config/agents/*.md config.

Right now subagents seem to always inherit their reasoning level from their parent session with no obvious way to override it. It would be very useful if a parent agent on Medium/High could spawn subagents with Low for faster results (other way around could be useful sometimes as well).

With Gemini 3.8 Flash being the only model right now, subagents seem to just slow down tasks rather than speed them up. 3.7 flash never had this problem as it was a lot faster and better results in general.

Custom subagents | Google Antigravity Docs describes that you can use “inherit” “flash” or “pro” as an argument to the subagent creation. For example, in Randal's Workflow (this week, anyway) · GitHub I use a light model to run the orchestrator, and delegate the code reviews to a “pro” model when needed.

Same here, tried using the subagents workflow for an actual feature today and it’s significantly slower than running on the main context window.

Already using 3.8 Flash Medium

Hi @James_Frowen, @RandalSchwartz, and @jerrydev,

Thank you for raising this and sharing your workflow setups.

  1. Current Subagent Model & Thinking Level Configuration:

    • In Antigravity App (v2.24.0) and Antigravity CLI (v1.4.2), you can configure the model tier for a subagent, though there is not yet a separate parameter in invoke_subagent or .agents/agents/*.md / ~/.gemini/config/agents/*.md YAML frontmatter to independently override the thinking level (Low, Medium, High).
    • As the Custom subagents YAML frontmatter supports setting model to inherit (default), flash, or pro (for example, subagent: true and model: flash in .agents/agents/<name>.md), and the invoke_subagent tool also accepts a Model parameter to override the frontmatter setting for a specific invocation.
  2. Feature Request Noted:

    • Being able to run an orchestrator agent on High or Medium while spawning read/search subagents explicitly pinned to Low thinking effort (or vice versa) is a great workflow improvement. We have shared this feature request with our engineering team to support a dedicated thinking-level option in custom subagent frontmatter and invoke_subagent.
  3. Tips to Speed Up Subagent Execution Today :

    • Restrict the Subagent Toolset: In your custom agent .md file (Custom subagents), restrict the tools list to only the read-only tools needed (such as view_file, grep_search, and find_by_name) and instruct the subagent in its prompt to return concise findings directly without broad repository sweeps.
    • Use Subagents for Parallelizable Tasks: Because each subagent starts with an isolated context window (Invoking subagents), delegating a single sequential task requires the subagent to re-read files already in your main session’s context. Subagents provide the biggest speedup when multiple independent tasks are launched in parallel.

Hi @Rebbanpalli_Naveen
Thanks for the clarification. However, there is an important nuance regarding point 1 for individual / personal accounts:

Even though individual accounts have access to Pro models in our sessions, configuring the subagent model via YAML frontmatter (model: pro) or passing Model: “pro” in invoke_subagent currently does not take effect for subagents. Under the hood, subagent calls still end up being routed/redirected to Gemini 3.8 Flash regardless of what is specified.