i've been testing OpenGradient's SDK for a few days now. nothing serious. just spinning up models from the hub, running basic inference, getting a feel for how it actually works versus how it's described. the first thing that struck me was how smooth the simple stuff is. upload a model, run inference, cryptographic proof attached without extra steps. the "verifiable by default" promise holds up cleanly when the task is straightforward. but then i tried chaining multiple inferences together. agentic workflow. multi-step reasoning. the kind of thing you'd actually build if you were serious. and the friction showed up fast. TEE attestations added latency between steps. a proof failed on an edge case that worked fine in isolation. nothing broken, nothing dramatic, just overhead. the system is strongest at simple hosting. the complexity is where the trade-off lives. i don't think that's a flaw. early infrastructure always has a gradient. but it made me wonder how many builders hit that friction and quietly move on before the tooling catches up

#opg $OPG @OpenGradient