Split work with subagents
Parallel only when the work can stay independent
After this, you'll be able to split a Codex task into independent subagent jobs and keep approvals, sandbox, and merge decisions under control.
Before you start
Complete Add skills and plugins without bloat first.
The idea
Subagents help when the work is actually parallel. Codex can spawn specialized agents in parallel and collect their results. That is useful for codebase exploration, independent audits, or a multi-step feature plan with clean boundaries.
Here is the before and after: before, Codex is guessing from a loose request. After, you can split a Codex task into independent subagent jobs and keep approvals, sandbox, and merge decisions under control.
Now try it use the exercise prompt on one real repo task. Keep the output small enough to check before you accept the change.
You are ready when the Codex action, boundary, and proof all match the task.
Try it (14 min)
Watch out for
- Splitting tasks that edit the same file.
- Letting subagents invent separate API contracts.
- Approving a command without checking which subagent requested it.
- Using subagents to avoid making the plan clear.
Paste this into Claude
Decide whether to use Codex subagents: Task: [describe task] Independent parts: [list] Shared files or contracts: [list] Can some parts be read-only: [yes/no] Approval risk: [low, medium, high] Final merge owner: [you, Codex main thread, unknown] Return: 1. Use subagents or stay sequential. 2. Subagent briefs if yes. 3. Shared contract each subagent must respect. 4. Approval rule for inactive subagent prompts.
What a good response looks like
Use subagents for read-only exploration only. Agent A maps route owners. Agent B maps tests. Agent C checks related docs. Main thread owns edits after comparing findings. Approval rule: no write or network approval from subagents without main-thread review.
What good looks like
- Subagents are used only for independent parts
- Shared contracts are named
- Read-only exploration is preferred when the split is uncertain
- Approval prompts from subagents are tied back to their source thread
When this breaks
- Breaks when parallel edits conflict because the shared contract was missing.
- Breaks when an inactive subagent approval gets approved without understanding its task.
AI can help with this
Use Codex to help you you can split a Codex task into independent subagent jobs and keep approvals, sandbox, and merge decisions under control. Start with the exercise prompt and your real input. Ask for one draft, then check it against this proof: Subagents are used only for independent parts. Accept only the version you can verify yourself.
You can now
You can identify independent work
- You can name shared contracts
- You can keep approvals tied to the requesting subagent
- You can keep main-thread ownership of merge decisions
Key takeaways
Use subagents for independent evidence or isolated work. Stay sequential when the contract is still unclear.
- Subagents can run specialized work in parallel.
- They inherit the current sandbox policy.
- Parallel work needs clear boundaries.
- Main-thread review still owns the final decision.
Go deeper