99 karma · joined October 15, 2023
jk its trash
not saying this is the case here but it does feel a bit like wine tasting sometimes, everyone claims to be an expert that can taste a few tokens and tell you exactly what region and vineyard its from.
to me they still are the only lab that ships fantastic models that are easily portable to different agentic harnesses. I use my codex subscription 24/7 within opencode and wingman and haven't had any complaints in a long time.
If OpenAI cuts alternative harness support it will be a weird day trying to figure out what to do next, it's been so clearly the best bang for your buck (imo) for a while. maybe id finally have to give smaller models a try.
anything to avoid using the dogwater codex & claude code tuis.
anyways this seems like a nice cost improvement over GPT 6 Sol and I expect this will be my new daily driver.
i think the language primitives map well to the natural desire to use these tools across networks and in concurrent workloads.
another Go harness i keep up to date with is: https://github.com/boldsoftware/shelley (exe.dev team)
https://docs.wingman.actor/concepts/sessions/
I think there is still plenty of innovation to be had in the space. Right now everything is still pretty bundled to the file system where getting skills and things to work inside of like a virtual worker requires simulating a virtual file system (usually). I believe even in the latest opencode v2 versions it still requires that a session be bundled to a working directory, I try to avoid (some of) that in Wingman.
I agree.
> The age of humans comprehending things is coming to an end: our brains just won't have the capacity to make meaningful contributions to science, math, or technology.
I don't know if I see this being true for quite a while, if ever.
I think your approach of a unified SDK/data structure around a chat that adapts to different clients is a really strong solution that probably just gives you the best of both worlds.