Agentic wrappers

OpenAI released their SWE coding agent today - Codex.

Sam tells us that 2025 is the year of agents, and it’s reasonable to think that after the launch of Operator in Jan ‘25 and now Codex several more similar releases are coming this year and next.

Although I am on the whole unimpressed with Operator from a UX perspective, and while I expect Codex to both be significantly more impressive and, initially, simultaneously frustratingly limited in its practical applicability, I do believe that OpenAI will solve the problem and create useful and helpful and applicable agent. Just how broadly applicable and how autonomous and how much economic value it will extract over time - who knows - but I’m certain they will make a significant dent in the problem.

The hype for generative AI as the next computing platform has attracted a rush of investment in startups trying to capitalize and create businesses in their little corner of the economy or, alternatively as they tell their investors, devise the thin edge of the brilliant strategic wedge which will lead to -something something- take over the world. We see plenty of agent startups and while many of them are not “real” agents (whatever that means) that doesn’t mean folks aren’t trying.

Heck, even I built a few toy agents in my time. Fido v2, whereby the dog training journey was dynamically managed by a machine intelligence and algorithm working in tangents with one another, was legitimately an agentic system.. and that was all the way back in the summer of ‘23! SchedulingAI was another forray into the world of empowered machine intelligence.

I struggle to see the strategic rationale for building a commercially focused agent in 2025 from the perspective of the indie hacker or startup founder. In the same way that OpenAI will eventually grind down the helpful-because-they-are-specialized-GPT-Wrappers with their increasingly-generalized-intelligence-ChatGPT-which-is-also-way-more-convenient-and-multiuse, I expect that oai-model-agentic-wrappers will eventually be crushed by OpenAI - or some other own-party-solution hyperscaler. Arguably it will happen much faster in the agentic world due to the higher order of inference and therefore cost required both to develop and then deploy the solution.

Agentic systems have a few core prerequisites of which some include: an intelligence engine, a data environment, an actionability interface.

OpenAI makes available the machine intelligence and the startups believe that a focused product experience and world class sales org will compell buyers to give them access to their data inputs and control over their systems.

But they miss OpenAI’s Trojan Horse: the ChatGPT desktop app.

If you’ve used apps with ChatGPT you already have an intuitive sense for this. The desktop app is the beachhead which gives OpenAI access to read the contents of files on user machines, and with additional permissions ChatGPT can also edit the contents of files. They’re experimenting with computer control, with all the safeguards blah-blah-safety-stuff that we collectively profess to care about.

Um, Hello! Data environment and actionability permissions. ChatGPT’s desktop app is the agentic interface and, hey!, your entire digital life, professional and personal, is accessible by one desktop or another. How easy! And convenient! While ChatGPT is here reading my IDE codebase, might as well go ahead and have it plan my vacation as well.

(As an aside, can we please as an industry come up with a more interesting and compelling example of an agentic application than planning a vacation, jeez)

I suspect that at the end of this cycle, when the dust settles, we will learn the same lesson we’ve learned over and over again in tech: the winner isn’t necessarily decided by who has the best deep tech, although that doesn’t hurt, but rather who’s got the highest quality platform. My bet is that OpenAI shifts from focusing on having the most powerful models to having the most powerful all-purpose platform. And that’s because that’s where the economic value is to be had.

There’s some echo of the flying cars and 140 characters somewhere in there - we were promised a cancer curing intelligence and we got another tech platform - but, hey, value is value.

My bet would be ChatGPT, or some successor app, is the AI platform which OpenAI will use to bludgeon all the wrappers and subsume the plurality of AI economic value.

The counterpoint to this thesis is the analog of what happened to Zuck during mobile.

Facebook learned in the early 2010s the importance of owning the platform, as relying on Apple for distribution and to set the rules of the road preventing Facebook from scaling to the extent and the depth desired. Privacy is an illusion, Zuck told us. Yes, ok. “Privacy, shmairacy…” until Apple wacks privacy as a tablestakes cost of doing business.

No wonder Zuck is hungry to own the models that he thinks are core to owning the next great platform.

And you could very well see Apple - conveniently, coincidentally - change their T&Cs for MacOS apps to restrict app access and controls to the files and systems which ChatGPT relies on to be more than just a very helpful chatbot. Because privacy, obviously.

No wonder Sam is working on a next gen computing device.

We’re all minnows watching the giants claw at one another for hypothetical next gen platforms. What a world we live in.