Copilot Chat Extension Version: 0.43.0
VS Code Version: Version: 1.115.0 (Universal)
OS Version: macOS 26.4.1
Feature (e.g. agent/edit/ask mode): Agent mode launching a subagent
Selected model (e.g. GPT 4.1, Claude 3.7 Sonnet): Claude Sonnet 4.6 (probably is affected for all MS CoPilot models)
Logs: na
Steps to Reproduce:
- Ask to do something with a sub agent. Tell the orchestrator to use a specific model from Microstoft CoPilot (no third party involved here)
- Verify that the sub agent did not run with that model, it has only run with the same model launching the sub agent
___________ AI writeup and suggestion ____________
AI WriteUp for better understanding:
Summary
When the agent invokes a sub-agent via runSubagent, there is no way to specify which model that sub-agent should use. The sub-agent always inherits the model selected in the VS Code UI by the user. This prevents the agent from delegating tasks to a more suitable model (e.g. a cheaper/faster model for simple lookups, or a more capable model for complex reasoning steps).
Current Behaviour
Sub-agents are spawned using the same model that was selected by the user for the top-level conversation. There is no model parameter on the runSubagent tool. Prompting the agent to "use Claude Opus for this sub-agent" has no effect — the model is fixed at the session level.
Expected Behaviour
The runSubagent tool should accept an optional model parameter:
{
"prompt": "Analyse this codebase for security vulnerabilities...",
"agentName": "Explore",
"model": "claude-opus-4-5"
}When omitted, the sub-agent inherits the session model (preserving current behaviour). When specified, the sub-agent runs with the requested model.
Constraints / Scope
- This should only be supported for official GitHub Copilot models (i.e. models available in the Copilot model picker). Third-party or custom model endpoints should not be injectable through this path.
- The available model identifiers should be validated against the user's current Copilot subscription/model access, and the call should fail gracefully (with a clear error) if the requested model is not available.
- This is an opt-in capability — existing agent definitions that don't specify a model are unaffected.
Use Cases
- Use a fast/cheap model (e.g. GPT-4o mini) for high-volume read-only exploration sub-agents.
- Use a high-capability model (e.g. Claude Opus) for complex reasoning or security analysis sub-agents, independent of what the user has selected for the main conversation.
- Agent definitions (
.agent.md) could declare a preferred model for their specific task domain.
Impact
Without this capability, the agent cannot optimise cost/quality trade-offs across sub-tasks, and power users cannot build multi-model agentic workflows within VS Code Copilot.
Copilot Chat Extension Version: 0.43.0
VS Code Version: Version: 1.115.0 (Universal)
OS Version: macOS 26.4.1
Feature (e.g. agent/edit/ask mode): Agent mode launching a subagent
Selected model (e.g. GPT 4.1, Claude 3.7 Sonnet): Claude Sonnet 4.6 (probably is affected for all MS CoPilot models)
Logs: na
Steps to Reproduce:
___________ AI writeup and suggestion ____________
AI WriteUp for better understanding:
Summary
When the agent invokes a sub-agent via
runSubagent, there is no way to specify which model that sub-agent should use. The sub-agent always inherits the model selected in the VS Code UI by the user. This prevents the agent from delegating tasks to a more suitable model (e.g. a cheaper/faster model for simple lookups, or a more capable model for complex reasoning steps).Current Behaviour
Sub-agents are spawned using the same model that was selected by the user for the top-level conversation. There is no
modelparameter on therunSubagenttool. Prompting the agent to "use Claude Opus for this sub-agent" has no effect — the model is fixed at the session level.Expected Behaviour
The
runSubagenttool should accept an optionalmodelparameter:{ "prompt": "Analyse this codebase for security vulnerabilities...", "agentName": "Explore", "model": "claude-opus-4-5" }When omitted, the sub-agent inherits the session model (preserving current behaviour). When specified, the sub-agent runs with the requested model.
Constraints / Scope
Use Cases
.agent.md) could declare a preferred model for their specific task domain.Impact
Without this capability, the agent cannot optimise cost/quality trade-offs across sub-tasks, and power users cannot build multi-model agentic workflows within VS Code Copilot.