One brain, several hands

I plan in one place and let other coding agents do the implementation. The split sounds tidy. It only works if you're honest about which agent gets which job.

In plain terms, one model reads the task, plans it and decides what "done" means. Then the actual code writing goes to whichever host suits that kind of task. Claude does the planning. Codex and a couple of cheaper hosted models do a lot of the typing.

In August I audited my own routing. One hosted model was the only one its vendor had certified for agent work, so it was my default for that tier. I ran identical probes against it, back to back. The answers were 200, 200, 429, 429 and 404. Same request, five outcomes. Then I gave it three real implementation attempts. All three failed.

I tried a stopgap model from the same provider. It fell over on basic tool-call formatting as soon as the task grew past something small and tightly scoped.

Then I gave the same work to a different provider. Research, JSON drafting, shell patches. It completed every task on the first try, at better quality than either of the first two.

The lesson wasn't that provider B is better. It was that a certification label had stopped matching reality, and my routing was still trusting the label. I now treat that preference as a point-in-time finding. I'll re-test it when the first provider looks stable again.

Here's how routing works today. Every class of task has a primary host and two fallbacks. The table lives in one file that I maintain by hand. When I sat down to plan a new policy file for this, the audit found that the old one already called itself the single source of truth. A second file would have been the same duplication I tell my own code not to do. So I extended the first.

Spend gets the same treatment. The newest, most expensive tier is reserved for the hardest work. A session warns at 75 requests, asks for a checkpoint at 150 and stops at 225. Large cached contexts still eat capacity, even when the cache makes them cheap per call.

What I'd tell anyone building this: keep the planner fixed, measure the hands, and assume any host can disappear on a Tuesday. The brain is the part that shouldn't move. The hands are suppliers, and suppliers need a backup.