Codex subagents: clean fan-out without polluting context

Subagents quickly make it feel like you have a team.

The trap is that the team sometimes works on a problem that only needed one careful read. You get four answers, three confidence levels, duplicates, and you still have to integrate the whole thing mentally.

Clean fan-out starts with one question: are these investigations independent?

1. When it is worth it

Use subagents when branches of work can move without stepping on each other.

Good cases:

• compare three architecture approaches;
• inspect five independent folders;
• check security, tests, performance, and UX on the same PR;
• prepare editorial angles from multiple official sources.

Bad cases:

• fixing one function in front of you;
• writing one short article;
• doing research where each step depends on the previous result;
• delegating only because the task is fuzzy.

Parallelism does not fix poor framing.

2. Require a normalized output

The parent must be able to compare results. Otherwise you get different formats and lose the benefit.

Useful prompt:

Spawn 4 subagents.
Each subagent must answer with:
- main finding;
- evidence with paths or sources;
- risk;
- recommendation;
- confidence: low, medium, or high.
Wait for all results and consolidate.

This forces agents to produce actionable evidence, not mini-essays.

3. Keep the parent as decision-maker

A subagent explores. The parent decides.

Do not ask every agent to propose a complete change to the same file. You will create incompatible recommendations. Ask instead:

• “find the risks”;
• “list the affected files”;
• “check whether a constraint exists”;
• “propose an angle, not the final patch”.

The final patch should remain consolidated in the main session.

4. Limit depth

Recursive fan-out is tempting, but it explodes quickly. One agent spawning three agents, each spawning three more, is not a team. It is a bill and a tree you cannot review cleanly.

For a blog, one level is enough:

• parent: decides the topic;
• agents: research sources, angles, conflicts;
• parent: writes or chooses.

5. Checklist before spawning

Before using subagents, check:

• tasks are independent;
• each agent has a clear question;
• output is normalized;
• the parent keeps final decision;
• token cost is worth the time saved.

If one point is missing, stay single-agent.

Sources

• OpenAI Codex - Subagents
• OpenAI Codex - Best practices