[FEATURE]: Allow the LLM to create subagents at runtime via the task tool #47751

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

Problem

Subagents currently have to be defined before a session starts — config files, agent directories, etc. When the model hits a job none of the existing agents fit (e.g. "diff these two PDFs and report only semantic changes"), it either stretches a generic agent like general or has to stop and ask the user to write config and restart. That breaks the flow for one-off specialized work.

Proposal

Let the task tool accept an inline agent definition alongside the prompt, e.g.:

task(
description: "review pdfs",
prompt: "Compare a.pdf and b.pdf for meaningful changes",
runtime_agent: {
name: "pdf-reviewer",
description: "Specialized PDF review agent",
prompt: "You are a specialized PDF comparison agent...",
tools: { read: true, grep: true }
}
)

The agent would be registered in-memory in the agent service and spawned as a normal child session — no restart, no files written to disk. Rules I'd keep:

• runtime agents can't shadow native or config-defined agents (name collision = error)
• the tools map acts as a strict allowlist (everything denied, then the listed tools allowed)
• permissions derived the same way as for static subagents

Benefits

One-off specialist agents on demand, mid-conversation. Complements file-based agents (which stay the right choice for recurring/team conventions) with no file-system side effects from the LLM.

I have a working implementation (~250 lines in agent.ts + task.ts) with a test suite and a live e2e run against a server built from source. Happy to open a PR if there's interest.