Subagents
Some work splits cleanly into pieces that do not need each other. Subagents run those pieces at the same time and hand back one answer.
Codex can run agents in parallel and collect their results into a single response. The documentation names the cases it suits: exploring a codebase, and working through a multi-step plan where the steps are independent.
Current releases enable subagent workflows by default, and their activity shows up in the desktop app, the CLI and the IDE extension, so you can see what each one did.
Why it is not just faster
- Parallel where the work is parallel. Four searches at once rather than four in a row.
- Separate context. Each subagent does its own model and tool work, so what it read does not fill your session.
- A summary comes back, not the whole transcript, which is the point when the detail is noise.
The cost is the same as the benefit: you get the conclusion rather than the work. When you wanted to read the work, run it in your own session instead.
Custom agents
In local clients you can define your own agents with different models and different instructions, so a search agent can run on a cheaper model than the one writing code. That is the same instinct as the profiles in lesson 13: match the setup to the job rather than using one setup for everything.
- Give Codex a task that is four independent lookups and watch how it splits it.
- Write down one job of yours that would suit a cheap, read-only custom agent.
Slow is fine. Stopping is the only problem.