First use case I'm putting to work is testing web apps as a user. Although it seems like this could be a token burner. Saving and mostly replaying might be nice to have.
I don’t think clicking buttons on a Mac is a particularly scary barrier. It’s not anymore scary then running an LLM in agent mode with a very large number of auto-approve programs and walking away for 15 minutes.
There seems a fair enthusiasm in the UI of these to hide code from coders. Like the prompt interaction is the true source and the actual code is some sort of annoying intermediate runtime inconvenience to cover up. I get that productivity can be improved with a lot of this for non developers, just not sure using 'code' as the term is the right one or not.
I think this would work much better if there were constraints in place, a software stack clearly separating different concerns - e.g. you just ask AI to write business logic while you already have data sources, auth, etc, configured.
But that's not how popular, modern software stacks work. They are like "you can do anything, anything at all!".
Consider Visual Basic for Applications - normally your code is together with data in one document, which you can send to colleague. It can be easily shared, there's nothing to set up, etc.
That's not true for JS, Python, Java, etc - you need to install libraries, you need to explicitly provide data, etc. Software industry as a whole embraced complexity because devs are paid to deal with complexity.
Now AI has to use same software stacks as the rest of the industry, making software fragile, requiring continuous maintenance, etc. VBA code which doesn't use any arcane features would require no maintenance and can work for decades.
So my guess is that the bottleneck might be neither models nor harness/wrapper - but overall software flimsiness and poor architectural decisions
I think the intent is more "we won't need coders" ... the real goal is to get to the point where Product Managers can just write specs and a working product comes out the other end.
These people HATE that developers have been necessary and highly paid and, in their view, prima donnas. I think most of the people running these companies actually despise developers.
My monthly subscription for Claude is up in a week, is there any compelling reason to switch to Codex (for coding/bug fixing of low/medium difficulty apps)? Or is it pretty much a wash at this point?
I've been switching between both depending on which one is having a good week — and that's the honest answer for most people right now.
But the real issue I ran into wasn't which model is better. It's that every time I switched, I lost weeks of accumulated context. The AI didn't know my project's conventions anymore, didn't remember the architecture decisions, didn't know what was tried and rejected.
What helped me was separating the project context from the tool. Keep the conventions, rules, and decisions in plain files in the repo. Both Claude Code and Codex can read them at session start. Then the question becomes "which model is sharper this week" instead of "can I afford to lose my context."
The answer to your question: it's mostly a wash on capability. The real cost of switching is the context you don't realize you're rebuilding.
My current expectation is that the Cowork/Codex set of "professional agents" for non-technical users will be one of the most important and fastest growing product categories of all time, so far.
i.e. agents for knowledge workers who are not software engineers
A few thoughts and questions:
1. I expect that this set of products will be extremely disruptive to many software businesses. It's like when a new VP joins a company, they often rip and replace some of the software vendors with their personal favorites. Well, most software was designed for human users. Now, peoples' agents will use software for them. Agents have different needs for software than humans do. Some they'll need more of, much they'll no longer need at all. What will this result in? It feels like a much swifter and more significant version of Google taking excerpts/summaries from webpages and putting it at the top of search results and taking away visits and ad revenue from sites.
2. I've tried dozens of products in this space. For most, onboarding is confusing, then the user gets dropped into a blank space, usage limits are uncompetitive compared to the subsidized tokens offered by OpenAI/Anthropic, etc. It's a tough space to compete in, but also clearly going to be a massive market. I'm expecting big investment from Microsoft, Google etc in this segment.
3. How will startups in this space compete against labs who can train models to fit their products?
4. Eventually will the UI/interface be generated/personalized for the user, by the model? Presumably. Harnesses get eaten by model-generated harnesses?
Products I've tried: ai browsers like dia, comet, claude for chrome, atlas, and dex; claw products like openclaw, kimi claw, klaus, viktor, duet, atris; automation things like tasklet and lindy; code agents like devin, claude code, cursor, codex; desktop automation tools like vercept, nox, liminary, logical, and raycast; and email products like shortwave, cora and jace. And of course, Claude Cowork, Codex cli and app, and Claude Code cli and app.
Edit: Notes on trying the new Codex update
1. The permissions workflow is very slick
2. Background browser testing is nice and the shadow cursor is an interesting UI element. It did do some things in the foreground for me / take control of focus, a few times, though.
3. It would be nice if the apps had quick ways to demo their new features. My workflow was to ask an LLM to read the update page and ask it what new things I could test, and then to take those things and ask Codex to demo them to me, but it doesn't quite understand it's own new features well enough to invoke them (without quite a bit of steering)
4. I cannot get it to show me the in app browser
5. Generating image mockups of websites and then building them is nice
Maybe. The point is that in case of software it is fairly easy to verify if that what LLM produced is correct or not. Compiler checks syntax, we can write tests, there is whole infrastructure for checking if something works as expected. In addition, LLM are just text generating algorithms and software is all about text, so if LLM see 1 000 000 a CRUD example in Python, it can generate it easily, as we have a lot of code examples out there thanks to open source.
That's why LLMs shine in coding tasks. If you move to other parts of engineering, like architecture, construction or stuff like investment (there is no AI boom there, why?) where there is no so much source text available, tasks are not so repeatable like in software, or verification is much more complicated, then LLM-s are no longer that useful.
In software also I believe we will see soon that a competitive advantage have not those who adopted LLM, but those who did not. If you ask LLM what framework/language/approach use for a given task, contrary to what people think, LLM is not "thinking", it just generates text answer on the base of what it was trained on, so you will get again and again same most popular frameworks/langs/approaches suggested, even if there is something better, yet not that popular to get into model weights in a significant way.
I agree, and I think this extends to programming too. A lot of of software practices are built on the expectation humans are writing, reviewing and shipping code with that quickly becoming the case, processes, practices and even programming languages themselves will evolve to what agents need, rather than humans.
a version of Conway's law aimed specifically at agentic communication rather than human.
I keep seeing sentiment like this. I work for a relatively cutting edge healthcare enterprise as a sysadmin, and we've only just been given access to copilot chat. I don't think we're going to be having agents doing work for us any time soon.
I still think we're several "my agent sent an inappropriate email to all my contacts" away from people figuring out proper security controls for these things
Codex is my favorite UX for anything as it edits the files and I can use the proper tooling to adjust and test stuff, so in my experience it was already able to do everything. However lately the limits seem to have got extremely tight, I keep spending out the daily limits way too quickly. The weekly limits are also often spent out early so I switch to Claude or Gemini or something.
I imagine the generous limit we felt were just from the 2x codex was offerring. I also felt the regression, and only recently remembered they had this.
I wish Codex App was open source. I like it, but there are always a bunch of little paper cuts that, if you were using codex cli, you could have easily diagnosed and filed an issue. Now, the issues in the codex repo is slowly becoming claude codish – ie a drawer for people's feelings with nothing concrete to point to.
I can't help but see some things as a solution in search of a problem every time I see these examples illustrating toy projects. Cloud Tic Tac Toe? Seriously?
I swear OpenAI has 2-3 unannounced releases ready to go at any time just so they can steal some thunder from their competitors when they announce something
Side note: I really wish there was an expectation that TUI apps implemented accessibility APIs.
Sure we can read the characters in the screen. But accessibility information is structured usually. TUI apps are going to be far less interesting & capable without accessibility built-in.
They (and other AI players) have been using WAU over DAU for all their metrics, and many have questioned why. But if you look at other data sources of AI adoption, the reason is clear: Even while 56% of Americans now "regularly" use GenAI on a weekly basis, a much smaller percentage 10 - 14% use it on a daily basis. Here's one source but others had similar numbers: https://www.genaiadoptiontracker.com/
56% is much more impressive than 14%.
This may look bad until you consider that all of them are already desperately strapped for compute. I think the lower DAU is due to a combination of that and people still figuring out how to use AI.
I'm sure it's been said before, but more and more our development work is encroaching on personal compute space. Even for personal projects.
A reminder to me to air gap those to spaces with separate hardware [:cringe:]
114 comments
[ 3.6 ms ] story [ 74.8 ms ] threadI'm still paranoid about keeping things securely sandboxed.
But that's not how popular, modern software stacks work. They are like "you can do anything, anything at all!".
Consider Visual Basic for Applications - normally your code is together with data in one document, which you can send to colleague. It can be easily shared, there's nothing to set up, etc.
That's not true for JS, Python, Java, etc - you need to install libraries, you need to explicitly provide data, etc. Software industry as a whole embraced complexity because devs are paid to deal with complexity.
Now AI has to use same software stacks as the rest of the industry, making software fragile, requiring continuous maintenance, etc. VBA code which doesn't use any arcane features would require no maintenance and can work for decades.
So my guess is that the bottleneck might be neither models nor harness/wrapper - but overall software flimsiness and poor architectural decisions
These people HATE that developers have been necessary and highly paid and, in their view, prima donnas. I think most of the people running these companies actually despise developers.
But the real issue I ran into wasn't which model is better. It's that every time I switched, I lost weeks of accumulated context. The AI didn't know my project's conventions anymore, didn't remember the architecture decisions, didn't know what was tried and rejected.
What helped me was separating the project context from the tool. Keep the conventions, rules, and decisions in plain files in the repo. Both Claude Code and Codex can read them at session start. Then the question becomes "which model is sharper this week" instead of "can I afford to lose my context."
The answer to your question: it's mostly a wash on capability. The real cost of switching is the context you don't realize you're rebuilding.
i.e. agents for knowledge workers who are not software engineers
A few thoughts and questions:
1. I expect that this set of products will be extremely disruptive to many software businesses. It's like when a new VP joins a company, they often rip and replace some of the software vendors with their personal favorites. Well, most software was designed for human users. Now, peoples' agents will use software for them. Agents have different needs for software than humans do. Some they'll need more of, much they'll no longer need at all. What will this result in? It feels like a much swifter and more significant version of Google taking excerpts/summaries from webpages and putting it at the top of search results and taking away visits and ad revenue from sites.
2. I've tried dozens of products in this space. For most, onboarding is confusing, then the user gets dropped into a blank space, usage limits are uncompetitive compared to the subsidized tokens offered by OpenAI/Anthropic, etc. It's a tough space to compete in, but also clearly going to be a massive market. I'm expecting big investment from Microsoft, Google etc in this segment.
3. How will startups in this space compete against labs who can train models to fit their products?
4. Eventually will the UI/interface be generated/personalized for the user, by the model? Presumably. Harnesses get eaten by model-generated harnesses?
A few more thoughts collected here: https://chrisbarber.co/professional-agents/
Products I've tried: ai browsers like dia, comet, claude for chrome, atlas, and dex; claw products like openclaw, kimi claw, klaus, viktor, duet, atris; automation things like tasklet and lindy; code agents like devin, claude code, cursor, codex; desktop automation tools like vercept, nox, liminary, logical, and raycast; and email products like shortwave, cora and jace. And of course, Claude Cowork, Codex cli and app, and Claude Code cli and app.
Edit: Notes on trying the new Codex update
1. The permissions workflow is very slick
2. Background browser testing is nice and the shadow cursor is an interesting UI element. It did do some things in the foreground for me / take control of focus, a few times, though.
3. It would be nice if the apps had quick ways to demo their new features. My workflow was to ask an LLM to read the update page and ask it what new things I could test, and then to take those things and ask Codex to demo them to me, but it doesn't quite understand it's own new features well enough to invoke them (without quite a bit of steering)
4. I cannot get it to show me the in app browser
5. Generating image mockups of websites and then building them is nice
That's why LLMs shine in coding tasks. If you move to other parts of engineering, like architecture, construction or stuff like investment (there is no AI boom there, why?) where there is no so much source text available, tasks are not so repeatable like in software, or verification is much more complicated, then LLM-s are no longer that useful.
In software also I believe we will see soon that a competitive advantage have not those who adopted LLM, but those who did not. If you ask LLM what framework/language/approach use for a given task, contrary to what people think, LLM is not "thinking", it just generates text answer on the base of what it was trained on, so you will get again and again same most popular frameworks/langs/approaches suggested, even if there is something better, yet not that popular to get into model weights in a significant way.
Interesting times, anyway.
a version of Conway's law aimed specifically at agentic communication rather than human.
https://github.com/openai/codex/issues/2847
Does anyone know of a good option that works on Wayland Linux?
I swear OpenAI has 2-3 unannounced releases ready to go at any time just so they can steal some thunder from their competitors when they announce something
</tin foil hat>
Faster LLMs will be here by next year.
Sure we can read the characters in the screen. But accessibility information is structured usually. TUI apps are going to be far less interesting & capable without accessibility built-in.
It is instructive that they decided to go with weekly active users as a metric, rather than daily active users.
56% is much more impressive than 14%.
This may look bad until you consider that all of them are already desperately strapped for compute. I think the lower DAU is due to a combination of that and people still figuring out how to use AI.
Bunch of startups need to pivot today after this announcement including mine