Running multiple AI agents together in a self-coordinating Codex loop
The approach splits one task across several AI agents that communicate and coordinate with each other to get it done, without using Codex's built-in '/goal' feature. Roles are assigned to different models at different levels: planning uses a model set to 'xhigh' effort, backend work uses another model at 'high' effort, and frontend work plus use yet another model at 'high' effort. A separate model handles fixes flagged by reviewers, another acts as the that keeps the whole loop moving, and another checks whether the loop is still healthy.
The model used as needs more hands-on steering at first, but once it understands how the loop works, things run smoothly. Codex's response speed has been slow lately overall, and the model used for coordination in particular takes almost twice as long as the others to finish the same task. Because of that slowdown, some of these loops have been running nonstop for over a week, pausing only when usage runs out.
The model chosen for backend work is picked because it judges things more objectively than the newer one and burns less usage on token-heavy coding, while a different model is used for frontend work because it's noticeably better at that than the .
Key points
- Splits one task across multiple AI agents that self-coordinate rather than relying on a single model
- Assigns different models and levels to planning, backend, frontend, , fixes, coordination, and health-checking
- The -role model needs more steering at first but stabilizes once it understands the loop
- Some models take almost twice as long as others to finish the same task, creating a real speed gap
- Loops can run nonstop for over a week, stopping only when usage is exhausted