Add opt-in autonomous delegation setting for subagent-heavy workflows
What variant of Codex are you using?
CLI
What feature would you like to see?
I use Codex in workflow/repo setups that rely heavily on bounded subagent use. Today the built-in spawn policy requires me to explicitly restate delegation permission in the prompt (for example: "use subagents when needed" or "delegate this") even when that is already the intended workflow for the session.
I'd like a first-class opt-in setting, profile, or repo-level mode that allows Codex to autonomously use bounded subagent delegation when appropriate, without needing per-turn authorization.
Important scope boundary:
- I am not asking to enable subagents by default for everyone.
- I am not asking for unconstrained fan-out.
- I am asking for a user-controlled opt-in mode for subagent-heavy workflows.
A plausible shape:
explicit(current behavior): only delegate when the user explicitly asksworkflow: allow bounded delegation when repo/workflow instructions explicitly call for itauto: allow bounded sidecar delegation when useful
Guardrails I would still want:
- conservative max depth / fanout
- no duplicate parent-thread work after a task has been delegated
- clear visibility in the UI/history when Codex delegated and why
- conservative handling around approvals and other security-sensitive actions
Related context:
- #16996 policy conflicts with repo/workflow instructions
- #18105 orchestrator/delegate mode
- this request is narrower: expose an explicit user-facing setting so users who intentionally want autonomous delegation do not have to keep re-authorizing it every turn
Additional information
This is mostly a UX/product request. The current behavior feels like a tax for agent-heavy workflows: the capability exists, but I have to keep re-stating permission for it in natural language. An opt-in setting would preserve the conservative default while making multi-agent workflows much smoother for users who actually want them.
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗