Codex Subagents - Parallel custom agents for complex tasks

Hi everyone!

Codex just leveled up with Subagents — you can now spawn specialized parallel agents for complex tasks like PR review or multi-step features. Each subagent gets its own instructions, model, and tools, and Codex merges everything back cleanly.

Over the last week I used Codex to design, debug, and do embedded work for a new device prototype, and the speed honestly shocked me. This feature makes that whole experience feel even more serious. Now I can have one agent map, one review, and one check docs, and the main thread stays much cleaner instead of drowning in logs and side quests.

It really feels like OpenAI is going all in on the coding lane right now. This puts some real pressure on @Claude Code. And Goolgle, @Google Antigravity alone probably is not enough :)

Spawning specialized parallel agents for complex coding tasks is the right evolution for Codex — splitting a large problem into concurrent subagents that each handle a focused piece mirrors how experienced engineering teams actually decompose work, and doing it in parallel rather than sequentially should dramatically cut time-to-completion on multi-file refactors and complex feature builds. How do subagents coordinate when their changes overlap — is there a central orchestrator that detects conflicting edits across parallel workers, or do they operate on isolated branches that get merged at the end?

The concept looks solid.

What was the hardest part of building it?

Really interesting direction.

What stands out here is that this feels like more than just “parallel agents” as a feature.

Once the main agent starts spawning specialized workers, routing tasks, waiting on results, and consolidating outputs, the system begins to behave less like a single coding assistant and more like a coordination runtime for agent teams.

That feels like an important shift.

The value is no longer just better execution inside one context window, but structured delegation across multiple cognitive threads without everything collapsing into context rot.

Curious how you think about this long term.

Do you see subagents remaining a productivity feature inside Codex, or becoming a more foundational coordination layer for multi-agent software workflows?

This is very cool

So how do plugins like compound engineering work going forward?

Gonna try this out. Codex 5.4 is already awesome and this seems like a way to supercharge it.

Subagents feel like the first real step toward structured parallelism in coding workflows, not just bigger context windows. Splitting roles (explore, implement, review) while keeping the main thread clean solves a real bottleneck.

Avoiding context rot alone is a big deal for multi-step tasks.

Curious how coordination works in practice: when subagents produce conflicting changes, is there a central arbitration layer or does everything rely on post-merge reconciliation?

I use Cursor,Cluade Code, Codex at the same time.

Normally, I can tell to spin out 3/5 agents to solve GitHub issues within Cursor and Claude Code.

Codex can open 3/5 threads to parallel work but doesn't automatically spin out subagents.

Will try today.