Follow-up from talking with @roblourens.

• Enable by default
• Rename execute_prompt to an agent-focused name runSubagent.
• 💪 System instructions to use the tool for running read-only research tasks
• 💪 Ability to set an agent id (nee chatmode) for the given prompt (requires that the orchestrating agent knows

Why: The main benefit spawning isolated agent loops is that they can conduct extensive research in their own context window and return a concise summary to the main agent, preserving valuable space in the main conversation's context before it needs to summarize. The main use cases are exploration tasks, research, and independent work that can be clearly scoped.

Why not task tool, like Claude Code or Cline? VS Code already has several task primitives that this would clash with.

cc @aeschli