Problem
Currently, when using the Task tool to invoke subagents, the model specified in the Agent configuration file is ignored. All subagents use the main conversation’s model
regardless of their configuration.
Current Behavior
# agents/code-explorer.md
---
name: code-explorer
model: mmodel # MiniMax-M2.5 (0.2x cost)
---
When calling Task({subagent_type: "code-explorer", ...}), the agent uses the main conversation's model instead of mmodel.
Expected Behavior
The Task tool should respect the model field in Agent configuration, allowing different agents to use different models.
Use Case: Cost Optimization
Task Type │ Model │ Cost │ Benefit
───────────────────────────────────────────┼──────────────────────────────────────────┼──────────────────────────────────────────┼──────────────────────────────────────────
Code Exploration │ MiniMax-M2.5 │ 0.2x │ Fast, cheap
Architecture Design │ GLM-5 │ 0.5x │ Deep reasoning
Code Generation │ Qwen3.5-Plus │ 0.2x │ Programming expert
Documentation │ Kimi-K2.5 │ 0.3x │ Long text
Benefits
1. Cost Optimization: Use cheaper models for simple tasks
2. Specialization: Match model strengths to task types
3. Parallel Efficiency: Multiple agents with different models can work concurrently
Suggestion
Add an optional model parameter to the Task tool, or respect the Agent configuration's model field.
--------
Related Project: UltraWork - Multi-agent orchestration system https://github.com/paskaa/ultrawork-agent