Most products treat the agent as a feature: there is a real backend doing the work, and somewhere off to the side a model answers questions about it. I wanted to try the inversion — put OpenClaw underneath the web app and let it be the thing that actually handles the request. The frontend just sends intent and renders whatever comes back.
It works better than it has any right to, and it fails in genuinely new ways. Latency becomes a design problem rather than an engineering one. Idempotency and permissions matter far more than they do in a normal API, because the backend can decide to do something you never wrote down. I am writing up the architecture properly, but the short version: this pattern is early, it is real, and the constraints are not the ones you expect.