Summary
CustomAgentConfig currently has no model field. Sub-agents spawned via task() always inherit the session-level model. There's no way to run a sub-agent on a different (e.g. cheaper/faster) model.
Use case
We run an orchestrator where the main agent uses claude-opus-4.6 for complex reasoning, but delegates research tasks to a sub-agent that only calls MCP tools (code search, docs lookup). This sub-agent doesn't need a frontier model — claude-haiku-4.5 would be sufficient and 5-10x cheaper.
Currently, the sub-agent always runs on the session model, so every task() delegation costs the same as a main-agent turn.
Proposal
Add an optional model field to CustomAgentConfig:
interface CustomAgentConfig {
name: string
description?: string
prompt?: string
tools?: string[]
mcpServers?: Record<string, McpServerConfig>
infer?: boolean
model?: string // ← new
}When set, the sub-agent's LLM calls should use this model instead of the session-level one.
Environment
@github/copilot-sdk 0.1.33-preview.0- Copilot CLI 1.0.5
- Node.js / TypeScript
Summary
CustomAgentConfigcurrently has nomodelfield. Sub-agents spawned viatask()always inherit the session-level model. There's no way to run a sub-agent on a different (e.g. cheaper/faster) model.Use case
We run an orchestrator where the main agent uses
claude-opus-4.6for complex reasoning, but delegates research tasks to a sub-agent that only calls MCP tools (code search, docs lookup). This sub-agent doesn't need a frontier model —claude-haiku-4.5would be sufficient and 5-10x cheaper.Currently, the sub-agent always runs on the session model, so every
task()delegation costs the same as a main-agent turn.Proposal
Add an optional
modelfield toCustomAgentConfig:When set, the sub-agent's LLM calls should use this model instead of the session-level one.
Environment
@github/copilot-sdk0.1.33-preview.0