Feature request: native runSubagent / inline sub-agent tool

[tore-unumed]

Summary

Request native support for an inline sub-agent tool (similar to VS Code Copilot's runSubagent) that allows a workflow agent to:

1. Spawn an isolated sub-agent with a specific prompt and limited context
2. Wait synchronously for its completion (blocking, not fire-and-forget)
3. Receive the sub-agent's output as a return value in the parent agent's context

Motivation

Complex workflows benefit from parallelizing research across large codebases. For example, a specs-reviewer agent needs to search 12+ repositories for relevant patterns — doing this serially is slow and burns tokens on accumulated context.

In VS Code Copilot, runSubagent solves this by spawning stateless child agents that return a single result. Each sub-agent has isolated context, preventing token bloat in the parent.

Current limitations

• dispatch-workflow: Fire-and-forget — no return values, no waiting, cannot coalesce results
• workflow_call: Static YAML — cannot dynamically determine number/assignment of sub-tasks at runtime
• MCP Scripts: Run on the runner host but cannot invoke the Copilot CLI because auth tokens are proxy-held and scrubbed from the agent environment
• Confirmed: runSubagent tool is not available in the gh-aw Copilot CLI environment (tested in runs 22843694716 and 22843793775)

Proposed behavior

// Pseudocode usage inside a workflow agent
const result = runSubagent({
  prompt: "Search repos/ehr/ and repos/erp/ for all GraphQL resolvers related to billing. Return file paths and brief descriptions.",
  description: "Search billing resolvers",
  // Optional: restrict context (directories, tools, etc.)
});
// result is a string with the sub-agent's findings

Key properties:

• Synchronous: Parent blocks until sub-agent completes
• Isolated context: Sub-agent starts fresh, does not see parent's accumulated context
• Returns output: Sub-agent's final message becomes the return value
• Respects workflow config: Inherits network rules, MCP servers, permissions from parent workflow
• Token budget: Sub-agent has its own token/continuation budget

Alternatives considered

1. Serial execution: Works but is 3-5x slower and accumulates unnecessary context
2. dispatch-workflow fan-out with collector: Two-phase (dispatcher → workers → collector) pattern works but is complex, requires artifact/issue/comment coordination, and adds significant latency
3. MCP Script wrapping Copilot CLI: Blocked by the auth proxy architecture that (correctly) isolates API tokens

Environment

• gh-aw CLI v0.56.2
• Engine: copilot / claude-sonnet-4

[pelikhan]

This is a request for copilot-cli.