@deanwball pretty clear that Open AI intends Codex to absorb all of the productized utility of open claw. They don't expect you to use the chat app they expect you to use Codex gui for everything on your computer. (They are right, of course, as you and I both know)
Something about AI adoption that really surprises me is the number of non-technical people who are attracted to open-source agents like OpenClaw and Hermes Agent.
These agents are billed as somehow having a broader surface area than eg Claude Code or Codex. But the precise thing that caused me to call Opus 4.5 in Claude Code AGI back in November was that it was clear the agent had become good enough to achieve full computer use. That means the models of 6-8 months ago could use a computer to do a very large fraction of the tasks a human could do; that was enough for me to say, “AGI.” The rest would be better UI, easier integration with agents and the web/software tools, etc. The term “coding agent” was always a misnomer; if you knew how to ask the right way, the models of late 2025 could do anything.
It is fascinating, and I suppose very cool, that so many “normie” consumers are getting into agents by way of these open-source solutions. They permit a degree of data sovereignty, model choice, etc. that the proprietary agents do not.
Yet I’m not sure those things—which appeal principally to nerds—are what attract the normie group to the open-source agents (isn’t it nice we can use “open source” in the old-fashioned way with the agent products, as opposed to wrestling with “open-weight” and blah blah? I have always suspected that system-level stuff would be the more strategic thing to make OSS rather than the model parameters. Anyway, I digress).
I think instead the normies just like the form factor of the open-source coding agents more, which definitely lean into “agenticness” and “it can do anything!” more clearly than do the proprietary agents. You can just “get going,” as a novice, a little more cleanly than you can with the “coding agents” like Claude Code/Codex, which do assume some basic measure of technical competence (though it is totally possible to use those as a novice too).
The somewhat ironic result is that I see total novices—people who truly could not tell you what CLI stands for—finding better product-market fit in open-source agents than in agents made by trillion-dollar AI labs.
I wonder if this will be temporary—that, as the models improve, the labs will lean into that branding/form factor of OpenClaw, and with the advantages of vertical integration to boot. I kind of suspect so. Basic business school logic suggests that vertical integration wins out in the early phases of a product category.
But I do like this whole open agent adoption phenomenon among “normies,” on aesthetic grounds. It could also be a bulwark against lab concentration of power/anti-competitive strategy. People being able to switch models out within an open-source agent seems like a MUCH more durable “open-source AI” strategy than keeping our fingers crossed that someone, somewhere, finds it in their interest to give away multi-billion dollar artifacts for free.
I am somewhat concerned the open-source agents will have security issues, especially in the hands of novice users. But I doubt the marginal risk is THAT much greater than the marginal risk of people using proprietary agents, again especially novices.
