REA Teams running agents on real repos
Continue, Cline, or SDK users whose tools retry after a compile or a timeout and send the same shape again.
Where it helps
Repeats. Not first drafts.
The expensive failure is the agent that hits the same broken loop five times and pays the vendor for each pass. That is the job.
Who it is for
Leads who cannot explain the bill
The invoice grew. The model did not get smarter. The session repeated itself.
Platform owners
You already have a model vendor. You need a cap, a timeout, and a record of what burned a loop.
Where the help is partial
Every failure is a new structure → cache does not fire. The ceiling still applies. Unique work still costs vendor tokens. Local models get the same proxy contract; they do not infer faster because of REA.
Not for
Better prose
We do not improve the vendor’s writing. Cache freezes the first answer.
One-shot chat
A single question with no code and no retry is a thin wrap. Use the vendor directly.
A box with no model
If the machine cannot reach an OpenAI-shaped endpoint, a cache miss has nowhere to go.