Codex Subagent

This does one thing well: it spins up autonomous subagents that handle context-heavy tasks in their own token budget and return only the final result. The golden rule is simple. If a task plus its intermediate work would dump 3,000+ tokens into your parent agent's context, offload it to a subagent. The skill emphasizes intelligent prompting, meaning you need to give the subagent enough context, constraints, and direction upfront or it won't succeed. It's a straightforward pattern for keeping your main agent's context clean when you're dealing with research, file processing, or any task that generates mountains of intermediate output you don't actually need to see.

Install to Claude Code

npx -y skills add am-will/codex-skills --skill codex-subagent --agent claude-code