Gabba Builds · 11 November 2026
My AI team hit its usage limit. What happens next?
My OpenBot team hit a Claude usage limit, and we set an automatic Codex backup for two of the roles. The setting is configured and read back. We haven't observed the real switch and return yet.
The five-hour indicator showed 0% left. I don't have a reliable record of exactly which jobs were running when it ran out, so I'm not going to pin it on one task.
A single chat pausing is annoying. This was broader than that. Several agents share the same route, so when it runs dry more than one job can stall at once, and I'm the one who has to notice.
I use OpenBot as the place this AI team lives while I build the Gabba Works apps.
What happened when the indicator hit 0%
The first reaction is to want a backup, quickly. Fair enough. The questions underneath matter more than picking a second model name from a list.
Which work can another model actually pick up? When should an agent switch, rather than sit and wait? How do you get back to the preferred model once it's available again? A backup that can't answer those is only a setting.
I also don't want to pretend every task is a good candidate for an automatic hop. Some work is fine to keep moving. Some work I'd rather pause, because a different model may handle it differently, and I'll want to review the result.
What we configured
Two roles have Claude as their primary model and Codex as an automatic backup: the one that handles website and design work, and the one that does independent code and quality review. Those are the saved settings. They don't tell you which model answered any particular turn.
We checked the saved primary and backup settings, and that the automatic switch was turned on, for both roles. I wanted to know the backup was actually recorded, not only that someone meant to turn it on.
That's as far as the configuration work can take us. Being able to read the setting back shows it was saved. It doesn't show what the team does when a real limit hits.
Have we seen it actually switch?
No. We haven't observed the real switch and return. Nothing we have so far shows a real limit pushing an agent onto the backup, or the agent coming back once Claude is available again.
The indicator also showed a countdown of 41m beside the 0%. That's useful to know, but it's a timer on a screen. It doesn't prove when a particular agent will move back to its preferred model, so I won't put a return time on it.
A backup isn't automatically a like-for-like replacement either. Different models can handle the same task differently. For important work I'd want a suitable alternative, clear task context, and a review of the result. A second name in a setting isn't enough on its own.
If you're setting up a fallback of your own, this is the sequence I'd use:
- Identify the connected routes your setup actually supports, and choose a backup that is suitable for the work.
- Record the preferred route and the existing settings before you make a change.
- Configure the primary and the backup together using your system's supported controls, then read the saved settings back.
- Keep evidence of a normal limit event, the route used afterwards, and the return. Treat configuration and observed behavior as separate checks.
- Review important outputs for differences between models. Don't assume a timer guarantees an immediate switch back.
A prompt I've found useful: "Inspect my supported primary and backup routes and current saved settings. Explain what is configured and what evidence would show an actual limit-triggered switch and return. Don't change anything."
Configuration tells you a backup is in place. Only watching a real limit tells you whether it switches, and whether it comes back. I'm keeping those as two separate checks.
What's next
I'll watch an ordinary limit event, record which route was used afterwards, and check whether the agent returns when the preferred model is available again. Until that cycle has been seen, I'll keep calling this a configured plan, not a behavior I can already vouch for.
This team is how I build the Gabba Works tools, including PDF Desk, which signs and fills PDFs in your browser so files stay on your device. The usage-limit work is about keeping that kind of build moving when a model route pauses. It isn't a claim about the product itself.
The same team is what I'm trying to keep clear in Building an AI team without making a mess of the work.
If you've hit the same wall, I'd like to hear how you got around it. This stays a configured plan until I've watched a real limit all the way through.
Talk soon,
Gabba