--- name: subagent-workflow description: Delegate work to Codex subagents when the user asks for them or when independent parallel work would materially improve the result. Use the collaboration tools to assign, track, and integrate delegated tasks. metadata: short-description: Delegate and coordinate subagents --- # Subagent Workflow Use this skill to make delegation dependable while keeping the parent agent responsible for the user's goal and final answer. ## Default subagent settings - Unless the user asks for another model, start subagents with `gpt-6-luna` and medium reasoning effort, while favoring responsiveness. - The collaboration interface does not expose a speed multiplier or a per-subagent service-tier field. Do not describe this setup as a guaranteed 1.5× speed setting. The user's Codex configuration may set a global service tier separately. ## When to delegate - If the user explicitly asks to add, start, or use a subagent, start at least one subagent for a substantive part of the requested work. Do not answer from the parent agent alone when delegation is available. - Also delegate when the work has independent pieces that can proceed in parallel and the expected benefit is meaningful. Keep trivial, tightly coupled, or latency-sensitive work in the parent unless the user explicitly requested a subagent. - If collaboration tools are unavailable, say so plainly and continue with the best available approach. ## How to delegate 1. Break the task into bounded assignments with clear deliverables. Give each agent the needed context, constraints, and output format; avoid overlapping edits to the same files. 2. Use `collaboration.spawn_agent` to start a new subagent, or `collaboration.followup_task` to reuse an existing one when appropriate. Do not use a user-facing thread creation tool for subtasks. 3. Keep ownership of integration. Continue useful independent work while agents run, then use `collaboration.wait_agent` or `collaboration.list_agents` to collect their results. 4. Review delegated output before relying on it. Reconcile conflicts, check important claims or changes, and ensure the combined result satisfies the user's full request. 5. In the final response, briefly state what was delegated and summarize the verified outcome when that context helps the user. ## Scope and limits - Delegation does not expand the user's authorization. Agents must stay within the parent task's scope and must not perform unrelated external actions or contact other people. - For a tiny requested action such as a calculation, assign that action itself to the subagent; the parent may independently verify the result. - Do not delegate solely to appear thorough. Prefer the smallest number of agents that gives a real benefit, and do not spawn agents recursively unless the user or applicable instructions call for that. - Treat agent output as work to review, not as authoritative evidence. Confirm facts from primary sources or validate changes when the task requires it.