Codex subagents are almost great. So I finished the orchestration layer myself.

Short summary

The author uses Codex daily and finds native subagents promising but inconsistent across surfaces (desktop, CLI, tool-backed sessions). V1 is stable but lacks reliable per-child control over model, reasoning effort, and token budget; V2 is more ambitious but still behind a feature flag. The author built a custom orchestration layer to gain programmatic control over worker selection, model choice, and budget allocation.

• •Codex subagents work in parallel but capabilities vary across desktop, CLI, and tool-backed surfaces
• •V1 is stable but lacks reliable per-child model/effort/budget control; V2 is still experimental
• •Author built a custom orchestration layer for precise programmatic control over multi-agent workflows

Generated with AI, which can make mistakes.