[cjbarber]

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?

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

[postalcoder]

I agree with the sentiment but I think for normie agents to take off in the way that you expect, you're going to have to grant them with full access. But, by granting agents full access, you immediately turn the computer into an extremely adversarial device insofar as txt files become credible threat vectors.

For all the benefits that agents offer, they can be asymmetrically harmful. This is not a solved issue. That hurts growth. I don't disagree with your general points, though.

[avaer]

> for normie agents to take off in the way that you expect, you're going to have to grant them with full access

At this point it's a foregone conclusion this is what users will choose. It'll be like (lack of) privacy on the internet caused by the ad industrial complex, but much worse and much more invasive.

The threats are real, but it's just a product opportunity to these companies. OpenAI and friends will sell the poison (insecure computing) and the antidote (Mythos et all) and eat from both ends.

Anyone trying to stay safe will be on the gradient to a Stallmanesque monastic computing existence.

I don't want this, I just think it's going down that route.

[intended]

There was a recent Stanford study which showed that AI enthusiasts and experts and the normies had very different sentiment when it came to AI.

I think most people are going to say they dont want it. I mean, why would anyone want a tool that can screw up their bank account? What benefit does it gain them?

Theres lots of cases of great highly useful LLM tools, but the moment they scale up you get slammed by the risks that stick out all along the long tail of outcomes.

[ryandrake]

I agree, in general we are going to find that ultimately most employee end users don't want it. Assuming it actually makes you more productive. I mean, who the hell wants to be 10X more productive without a commensurate 10X compensation increase? You're just giving away that value to your employer.

On the other hand, entrepreneurs and managers are going to want it for their employees (and force it on them) for the above reason.

[TeMPOraL]

I want. If I get 10X more productive, I can unilaterally increase my compensation 10X by doing my stuff in 1 unit of time instead of 10 it took, and splitting the remaining 9 units of time into, say, 4 units of time doing more work, securing my position and setting myself up for promotion, and 5 units of time doing whatever the fuck I want. Not all compensation shows up in a bank account - working less, or under less stress, are also valuable.

Of course, such situation is only temporary - if I can suddenly be 10X productive, then so can everyone else, and then the baseline shifts so 10X is the new 1X.

[jbstack]

You want it, but then you closed by explaining exactly why you shouldn't want it. Plus, the new baseline isn't neutral (as in, everyone is the same again). If humans can now do 10x the work as before, the employer doesn't need the same number of humans to carry out its work. So the new baseline is actually "let's keep 1 employee and fire the other 9", unless the business can find a way to suddenly expand 10x so that it needs 10x as much work done.

[teiferer]

Your first paragraph is so short sighted that its message didn't even make it beyond the next one. It's a race to the bottom and your "doing whatever the fuck I want" will obviously never materialize.

The typical work week today is 40 hours. Just like it was 80 years ago. The typical worker is dramatically more productive than 80 years ago yet "doing whatever the fuck I want" time has not increased. Why would it? Employers don't need to pay such that 20 hour work weeks give you the same income. Because everybody around you is ok with working 40 hours.

This won't be different with AI, no matter if the overall effect is 1.1x or 10x or 100x productivity. Because it's not a technological problem but a sociological one.

[ryandrake]

Good point. My rant assumed that "10x productivity" meant 10x output in 1x time, rather than 1x output in 0.1x time. Only one of those are actually objectionable.

[hvb2]

> I mean, who the hell wants to be 10X more productive without a commensurate 10X compensation increase? You're just giving away that value to your employer.

Those are productivity increases that got our standard of living to where it is. Fewer people doing the same amount of work has, historically speaking, freed people from their current job, allowing them to work on something else.

It's that analogy of the horse, they used to be farm animals. Now, fewer of them are 'employed' but they're much nicer jobs. I'm not sure if the same is true for us this time around though as new jobs being created have increasingly been highly skilled which means the majority can't apply.

[drivebyhooting]

There was a long and great ravine of suffering between the advent of the Industrial Revolution and our time of bounty.

[Bombthecat]

Yep, all those artists, musicians, designers and coders will finally do something productive!

[yes_man]

If everyone becomes 10x more productive it won’t mean the companies cash flow 10x’s. Where value is loose there is competition, so in theory everyone should win. Unless nobody else can compete to capture that loose 10x value, in which case congratulations, you are now a unicorn.

Of course in reality in the short term what happens is companies lay off people to increase margins. Times will be tough for workers, and equity keeps gravitating towards those who already had it.

[King-Aaron]

Tasks have value because they take effort to complete.

If you remove the effort from those tasks, they will have no value.

10x the value of 0 is 0

[vovavili]

>Assuming it actually makes you more productive. I mean, who the hell wants to be 10X more productive without a commensurate 10X compensation increase?

Given sane working arrangements or at minimum presence of remote work, it would be a bit shortsighted not to want to get done with your work in a tenth amount of time. In the very least, you're competing for a promotion against less effective people, all while having more time for yourself. If not, you're building labor market skillset in an efficient way so you can hop to a better employer.

[procaryote]

It's interesting how differently people can think.

I couldn't imagine thinking "I'm gonna do this 0.1x as fast as I could, wasting my life away with pointless extra work, to spite my employer"

[teiferer]

> I mean, who the hell wants to be 10X more productive without a commensurate 10X compensation increase?

The person who realizes that everybody around them is bow at 10X and if they don't follow suit then they will soon be out of a job.

[teiferer]

> I think most people are going to say they dont want it. I mean, why would anyone want a tool that can screw up their bank account? What benefit does it gain them?

I'm not so sure. Matter of marketing and social pressure, big time.

Consider this: "Always-on pervasive google/fb/... login? I think most people are going to say they dont want it. I mean, why would anyone want a tool that would track their every move on the internet?" That could easily have been a statement 20 years ago. And look where we are.

[safety1st]

Their solution will be to push mandatory and nonconsensual updates to your devices which limit your device and your freedom in the name of security. Like Google is doing to Android in September. You will no longer be able to install "unverified" software on anything. To address prompt injection attacks they're probably working on an approach where your data all has to be in the cloud and subject to security scans. That's already basically the model for Google Workspace, Google Drive and Chromebooks.

The model will get full access to your data, but in the name of security, you will only be permitted to have data that is cloud-hosted; local storage will effectively just be cache.

The era of the general computer will end, and the products you purchased from these companies will be nonconsensually altered and limited.

I'm so glad I switched to Linux more than a decade ago. At least on the PC there will still be an open source ecosystem for a long time to come, it may have less features but I'm willing to accept that.

Knowing that they can change what you bought overnight with a single nonconsensual update, think very, very carefully about who you purchase all of your future technology from. Google's upcoming nonconsensual degradation of Android should be a lesson for everybody.

[WarmWash]

>Google's upcoming nonconsensual degradation of Android should be a lesson for everybody.

Google is almost certainly doing this because the iOS was not found to be a monopoly, while Andorid was. It came up in Google's appeal of the Epic case verdict, where they directly asked the judge about it. Turns out you can't be anti-competitive if you don't have [allow] any competitors.

[daveguy]

Nope. I'm still going to blame Google for their own actions. Nice try, though. I'm old enough to remember when Google pretended to take responsibility for not being evil. Even had it as their motto.

[shevy-java]

> I'm so glad I switched to Linux more than a decade ago. At least on the PC there will still be an open source ecosystem for a long time to come, it may have less features but I'm willing to accept that.

Wait until age verification is mandatory everywhere. :)

I can already see that happening, e. g. to access financial transactions or government apps, one needs to verify the id, and that will not work without age verification that can not be tampered with. So Linux will either submit to the same or be excluded.

(That free developers will be able to run Linux fine for much longer will also be true, but I guess they only care about catching the 95%, not the 5% linux users ... and 5% is a high guesstimate).

Edit: To clarify the above, one already had to provide personal data for financial transactions, of course, so a bank knows who is who, but the recent age verification go hand in hand with the attempt to get rid of vpn, and applications now make it a new standard to query the age of users, with the claim to "help protect kids". And some people buy into that rationale too. I don't, but I have seen many non-tech savvy people submit to that justification.

[soco]

There's always the zero knowledge proof tech alternative, but I don't have the feeling we are moving in that direction - it's not the most profitable business is it.

[duskdozer]

No, nor is it most amenable to mass surveillance.

[Springtime]

> It'll be like (lack of) privacy on the internet caused by the ad industrial complex, but much worse and much more invasive.

The concerning aspect is how others' content being scanned into systems don't have any knowledge or consent. Having private PII/files/code/emails/etc being read and/or accidentally shared by the agent online.

[hypfer]

> Anyone trying to stay safe will be on the gradient to a Stallmanesque monastic computing existence.

Honestly, it's alright.

Just think of what we could do with computers up until this point. We keep all those abilities.

And more, even, because the industry still keeps churning out new local LLMs. So you even gain more capabilities than right now. Just not at the rate of the bleeding edge.

Which is just like the Linux desktop, essentially. It's fine, really. There is no need to consume the bleeding edge. You will be fine.

[Forgeties79]

Definitely agree here. Made the swap to Linux a little over a year ago and the only reason I even have nice hardware is because I like gaming. But if I was cut off from everything tomorrow, the decades of stuff I have that I have not played will keep me very happy lol

[multjoy]

>Anyone trying to stay safe will be on the gradient to a Stallmanesque monastic computing existence.

As a proud neo-luddite, I'm watching the AI hype with grim amusement and I'll tell you hwhat, it doesn't look like a good time. Even putting to one side the planetary scale economic crash that is incoming, all the hypers seem to be on some sort of treadmill that is out of their control and it simply doesn't look like fun.

[petesergeant]

Do you think that avoidance is going to protect you from the fall-out?

[Forgeties79]

Everyone keeps saying how essential it all is yet a few years in and I still don’t see anything like the promised future of “everyone using them every day for everything.” Everyone’s just constantly talking (or stressing) about it.

We - including the companies - don’t know what the real “billion dollar application” of them is other than the unproven claim it makes everyone more productive in some general sense. When it doesn’t work people continue to say “it’s your fault not the tool’s.” Meanwhile investors are getting skittish and not one AI company is profitable yet. Companies that laid people off for LLM’s are regretting their decisions, leadership (and educators) is dealing with unvetted writing and having to waste their time cleaning it up, the list goes on. “Slop” is still a huge and growing problem.

LLM’s are here to stay, but IMO it’ll be more relevant in the long run than 3D printers yet less revolutionary than the internet. Everyone will touch them at various points but this whole-life, every-industry-disrupted integration still seems far fetched to me. Pricing is still a huge unsolved problem - everyone is still subsidized and despite gains in using fewer resources, it’s still too much to run these locally, even small models (not even getting into tooling and knowledge required to use them in a productive way).

When we zoom out and look at the whole picture, LLM’s have mostly made everyone’s online experience worse while the VC funded companies behind them are playing municipal and state governments’ for suckers a la Amazon getting so many cities to trip over each other giving away land and tax breaks, but far worse. Those are the biggest contributions so far aside from anecdotes from coders about “1000x productivity.” Again, I think they’re here to stay. But it’s called “AI hype” for a reason.

LLM’s have mostly been a problem creator IME rather than a “disruptor.” Never really seen “revolutionary technology” quite like it.

But hey, I’ll admit it’s useful to have a meh local model when I’m writing TTRPG stuff and have writer’s block. Though then I remember how it was trained, a whole other subject I haven’t even touched, so that kind of sucks too.

[multjoy]

Yes, mainly because I will continue to know the difference between a truth and a lie.

[elictronic]

2-3 news stories of people having bank accounts cleared and the product is dead on arrival.

[driverdan]

You'd think so but all the evidence so far points to the contrary. Most people seem perfectly happy to trade security and privacy for convenience.

[retinaros]

I dont see companies doing that. it can be business ending. only AI bros buying mac mini in 2026 to setup slop generated Claws would do that but a company doing that will for sure expose customer data.

[rurban]

Big companies are exposing customer data all the time, and they are doing all fine. The more criminal negligence, the richer.

[cjbarber]

> For all the benefits that agents offer, they can be asymmetrically harmful. This is not a solved issue.

Strongly agreed.

I saw a few people running these things with looser permissions than I do. e.g. one non-technical friend using claude cli, no sandbox, so I set them up with a sandbox etc.

And the people who were using Cowork already were mostly blind approving all requests without reading what it was asking.

The more powerful, the more dangerous, and vice versa.

[TeMPOraL]

> I saw a few people running these things with looser permissions than I do. e.g. one non-technical friend using claude cli, no sandbox, so I set them up with a sandbox etc.

People have different levels of safety-consciousness, but also different tolerances and threat models.

For example, I would hesitate running a Mythos-level model in YOLO mode with full control over my computer, but right now, for personal stuff, even figuring out WTF are sandboxes in Claude Code / Gemini CLI, much less setting them up, is too much hassle. What's the worst it can do without me noticing? Format the drive and upload some private data into pastebin? Much as I hate cloud and the proliferation of 2FA in every service, that alone means it can't actually do more to me than waste few hours of my life, as I reimage my desktop and restore OneDrive (in case of destructive changes that got synced up). These models are not yet good enough to empty my bank account in few minutes I'm not looking; everything else they can do quickly is reversible or inconsequential.

Now, I do look at things closely when working with agentic AI tools. But my threat model is limited to worrying about those few hours of my life. `rm -rf / --no-preserve-root` is an annoyance, not a danger.

(I accept that different contexts give different threat modeling. I would be more worried if I were doing businessy business stuff with all kinds of secret sauces, or was processing PII of my employer's customers, or lived in a country where it's easy to have all your money stolen if your CC number or SSN gets posted online.)

[planb]

How many of these threat vectors are just theoretical? Don’t use skills from random sources (just like don’t execute files from unknown sources). Don’t paste from untrusted sites (don’t click links on untrusted sites). Maybe there are fake documentation sites that the agent will search and have a prompt injected - but I haven’t heard of a single case where that happened. For now, the benefits outweigh the risk so much that I am willing to take it - and I think I have an almost complete knowledge of all the attack vectors.

[IanCal]

Systems have been caught out that review pull requests, that’s a simple and clear one. The more obvious to me for most people is anything you do that interacts with your email without an explicit approve list of emails to read.

[planb]

Yes, but none of this applies to the local codex agent that runs when I tell it to and has access to my computer. Like: „scan this folder of PDFs and create an excel file with all expenses. Then enter them into my tax software.“ This needs access to very sensitive data and involves a quite complex handling of data. But the only attack vector I see is someone injecting prompts into my invoice files.

[IanCal]

Which applies if you were to do this to invoices submitted to you, rather than ones you created, or if you have any way of user info getting into your invoices.

[teiferer]

The problem is that any data now becomes effectively an executable.

> I think I have an almost complete knowledge of all the attack vectors.

That's exactly the kind of hybris where the maximum danger lies.

[postalcoder]

i think you lack creativity. you could create a site that targets a very narrow niche, say an upper income school district. build some credibility, get highly ranked on google due to niche. post lunch menus with hidden embedded text.

the attack surface is so wide idk where to start.

[planb]

Why would my agent retrieve that lunch menu?

[thuuuomas]

Because it’s hooked up to a microphone in your kitchen & your kid is arguing with you about what lunch they want & they say “Hey [agent], what day is pizza day at [school]?”

[planb]

I’m not doing that. That would be like giving my child shell access to my system.

[boxedemp]

Funny joke,

But for real, obviously we all know people use agents to pick restaurants and that's a legit vector.

I agree it's not the biggest surface, but it's worth knowing imdo

[jasongi]

I cannot reconcile that growth for non-technical users is going to explode, when most utility from agents is via the ability to execute arbitrary code, generally in yolo mode, with the fact that almost all corporate IT departments do not give users the ability to install anything on their machine, let alone arbitrary code. Even developers at many companies are subject to this despite the productivity impacts.

The culture of corporate IT would need to change to allow it, and I just don't see it happening.

[Anvoker]

What about setting environments for normies that mitigate this problem? I don't know that you can do it on Windows, but Linux offers various tools for isolation where you can give full rights to an LLM and still be safe from certain classes of disaster.

Maybe this kind of isolation neuters the benefit you're thinking of, but I do believe some sort of solution could be reached.

[teiferer]

"Isolation" and "full rights" are mutually exclusive, contradictory properties.

[MrsPeaches]

This is me!

I’m semi-normie (MechEng with a bit of Matlab now working as a ceo).

I spend most of my day in Claude code but outputs are word docs, presentations, excel sheets, research etc.

I recently got it to plan a social media campaign and produce a ppt with key messaging and content calendar for the next year, then draft posts in Figma for the first 5 weeks of the campaign and then used a social media aggregator api to download images and schedule in posts.

In two hours I had a decent social media campaign planned and scheduled, something that would have taken 3-4 weeks if I had done it myself by hand.

I’ve vibe coded an interface to run multiple agents at once that have full access via apis and MCPs.

With a daily cron job it goes through my emails and meeting notes, finds tasks, plans execution, executes and then send me a message with a summary of what it has done.

Most knowledge work output is delivered as code (e.g. xml in word docs) so it shouldn’t be that that surprising that it can do all this!

[nonameiguess]

How does this obviate the need for software? In order for what you asked to be possible, Word, Excel, PowerPoint, and Figma all still need to exist and you need licenses for them.

If you can figure out the next step and say "Claude, go find me buyers and sell shit for me without using any pre-existing software," have at it. It can't be social media, I guess, since social media is software and Claude is supposed to get rid of software.

At a certain point, why do we even need computers? Can't we just call Claude's hotline and ask "Claude, please find a way to dump $40 million in cash into my living room. Don't put it in my bank account because banks use software."

[elAhmo]

It doesn't remove the need for software, but it greatly reduces the number of tools needed or doesn't mandate building custom tools that might not be viable due to very specific needs many users have.

OP gave a good example how their workflow was changed, you could argue there are tools that could've done that, but they managed to achieve their goals without them, have something that fits their workflow perfectly, is fine tuned in case of changes, and with a few other tools (Word, Excel, Figma) they can do all sorts of things which would've required a small team or far more (expensive) tools to execute.

To me that is a great example of non-developers using tools to enhance their workflows and with initiatives like from this topic, I can only see that increasing.

[TeMPOraL]

> How does this obviate the need for software?

It doesn't obviate the need for software, but it greatly devalues software products, as they become reduced to tool calls for LLMs.

This is good for users, because software products are defined by boundaries - borders drawn around the code to focus and package functionality, yes, but also to limit interoperability and create a sales channel (UX being the perfect marketing platform for captive audience).

After all, I don't usually want to play with Word, Excel, PowerPoint, and Figma - they're just standing between me and the artifact I want to create, so if I can get LLM to operate them for me, I don't have to deal with all the UX and marketing bullshit those products throw at me.

I mean, that's what I'd do if I could afford to hire a person to operate those tools for me. That, again, is the best mental model for LLMs - they're little people on a chip, cheaper to employ than actual people.

[DrScientist]

> I mean, that's what I'd do if I could afford to hire a person to operate those tools for me. That, again, is the best mental model for LLMs - they're little people on a chip, cheaper to employ than actual people.

Sounds like more of a threat to people than software then.

I get the point that if an agent could generate a presentation by directly writing to some open format with a free viewer then PowerPoint would be out of the picture.

However the tool has to be pretty close to 100% for that to work. If I have a presentation that's 90% there it's probably going to be a lot easier to finish it off manually in Powerpoint than try different variants of prompts. In which case I'll still need that Powerpoint license.

[drnick1]

> In order for what you asked to be possible, Word, Excel, PowerPoint, and Figma all still need to exist and you need licenses for them.

Or not. Besides, the better AI models can effortlessly generate Latex/Beamer, a far superior solution for typesetting and presentations. Anything than can be done in Excel can be done in Python. Those proprietary tools are a thing of the past, no one should use them anymore.

[Bombthecat]

And the value of those marketing campaigns is going to zero, since everyone is doing it. Even self employed people.

Pay for ads or you get lost in the mass of posts

[intended]

> 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 disagree. There is a major gap between awesome tech and market uptake.

At this point, the question is whether LLMs are going to be more useful than excel. AI enthusiasts are 100% sure that it’s already more useful than excel, but on the ground, non-technical views do not reflect that view.

All the interviews and real life interactions I have seen, indicate that a narrow band of non-technical experts gain durable benefits from AI.

GenAI is incredible for project starts. A 0 coding experience relative went from mockup to MVP webapp in 3 days, for something he just had an idea about.

GenAI is NOT great for what comes after a non-technical MVP. That webapp had enough issues that, if used at scale, would guarantee litigation.

Mileage varies entirely on whether the person building the tool has sufficient domain expertise to navigate the forest they find themselves in.

Experts constantly decide trade offs which novices don’t even realize matter. Something as innocuous as the placement of switches when you enter the room, can be made inconvenient.

[cjbarber]

> market uptake.

I think the market uptake of Claude Cowork is already massive.

[intended]

Estimated users are at 18-30 mn, and we are talking about non-technical users.

[bob1029]

> 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 agree this is going to be big. I threw a prototype of a domain-specific agent into the proverbial hornets' nest recently and it has altered the narrative about what might be possible.

The part that makes this powerful is that the LLM is the ultimate UI/UX. You don't need to spend much time developing user interfaces and testing them against customers. Everyone understands the affordances around something that looks like iMessage or WhatsApp. UI/UX development is often the most expensive part of software engineering. Figuring out how to intercept, normalize and expose the domain data is where all of the magic happens. This part is usually trivial by comparison. If most of the business lives in SQL databases, your job is basically done for you. A tool to list the databases and another tool to execute queries against them. That's basically it.

I think there is an emerging B2B/SaaS market here. There are businesses that want bespoke AI tools and don't have the discipline to deploy them in-house. I don't know if it is ever possible for OAI & friends to develop a "hyper" agent that can produce good outcomes here automatically. There are often people problems that make connecting the data sources tricky. Having a human consultant come in and make a case for why they need access to everything is probably more persuasive and likely to succeed.

[skydhash]

> The part that makes this powerful is that the LLM is the ultimate UI/UX.

I strongly doubt that. That’s like saying conversation is the ultimate way to convey information. But almost every human process has been changed to forms and structured reports. But we have decided that simple tools does not sell as well and we are trying to make workflow as complex as possible. LLM are more the ultimate tools to make things inefficient.

[duskdozer]

>The part that makes this powerful is that the LLM is the ultimate UI/UX

Seems pretty questionable to me. Describing things in natural language can be quite imprecise and verbose.

[voncheese]

>UI/UX development is often the most expensive part of software engineering.

I disagree with this as a blanket statement. At least in the tech world (i.e. tech companies that build technology products), UI/UX is often less expensive than the platform and infrastructure parts of the technology products, certainly at any tech that runs at scale.

[cjbarber]

> There are businesses that want bespoke AI tools and don't have the discipline to deploy them in-house. I don't know if it is ever possible for OAI & friends to develop a "hyper" agent that can produce good outcomes here automatically. There are often people problems that make connecting the data sources tricky. Having a human consultant come in and make a case for why they need access to everything is probably more persuasive and likely to succeed.

Sort of agreed, though I wonder if ai-deployed software eats most use cases, and human consultants for integration/deployment are more for the more niche or hard to reach ones.

[aerhardt]

I am starting to use Codex heavily on non-coding tasks. But I am realizing it works because I work and think like a programmer - everything is a file, every file and directory should have very precise responsibilities, versioning is controlled, etc. I don't know how quick all of this will take to spread to the general population.

[piokoch]

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.

Interesting times, anyway.

[jampekka]

LLMs nowadays make aggressive use of web search. Thus they don't answer only on the base of what they were trained on.

I don't think they are much more prone to using only the same popular frameworks, especially if you ask them to weigh for options.

[nazgulsenpai]

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.

[troupo]

> 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.

They won't.

Non-technical users expect a CEO's secretary from TV/movies: you do a vague request, the secretary does everything for you. LLMs cannot give you that by their own nature.

> And eventually will the UI/interface be generated/personalized for the user, by the model?

No. Please for the love of god actually go outside and talk to people outside of the tech bubble. People don't want "personalized interfaces that change every second based on the whims of an unknowable black box". They have plenty of that already.

[skydhash]

> Non-technical users expect a CEO's secretary from TV/movies: you do a vague request, the secretary does everything for you. LLMs cannot give you that by their own nature.

Most people are indifferent to computers. A computer to them is similar to the water pipeline or the electrical grid. It’s what makes some other stuff they want possible. And the interface they want to interact with should be as simple as possible and quite direct.

That is pretty much the 101 of UX. No deep interactions (a long list of steps), no DSL (even if visual), and no updates to the interfaces. That’s why people like their phone more than their desktops. Because the constraints have made the UX simpler, while current OS are trying to complicate things.

So Cowork/Codex would probably go where Siri is right now. Because they are not a simpler and consistent interface. They’ve only hidden all the controls behind one single point of entry. But the complexity still exists.

[noelsusman]

Just yesterday my non-technical spouse had to solve a moderately complex scheduling problem at work. She gave the various criteria and constraints to Claude and had a full solution within a few minutes, saving hours of work. It ended up requiring a few hundred lines of Python to implement a scheduling optimization algorithm. She only vaguely knows what Python is, but that didn't matter. She got what she needed.

For now she was only able to do that because I set up a modified version of my agentic coding setup on her computer and told her to give it a shot for more complex tasks. It won't be trivial, but I do think there's a big opportunity for whoever can translate the experience we're having with agentic coding to a non-technical audience.

[paganel]

There's no such big opportunity, as the number of programmers' spouses is quite limited. Again, and as the GP rightly suggested, some of the HN-ers here need to go and touch some normie grass, so to speak.

More to the point, nobody wants to be more efficient for the sake of being efficient, we all want to go to work, do our metaphorical 9 to 5 without consuming too much (intellectual and not only) energy, and then back home. In that regard AI is seen as an existential threat to that "lifestyle" and it will be treated as such by regular workers.

[w2df]

correct. you cant trust this place for realistic takes - I had a post re. financial stuff downvoted when a former Investment Banker chimed in to back me up.

Comical. Truly comical.

[troupo]

> Just yesterday my non-technical spouse

> It ended up requiring a few hundred lines of Python

And she knows those a hundred lines of python work correctly and give her correct result because in this instance Claude managed to produce a working result. What if it didn't? Would vague knowledge of Python have helped her?

> It won't be trivial, but I do think there's a big opportunity for whoever can translate the experience we're having with agentic coding to a non-technical audience.

Even though I agree with the sentiment, we've tried non-coding coding how many times now? Once every 5 years? Throwing LLMs into the mix won't help much when in the end you leave the end user hanging, debugging problems and hunting for solutions.

[zozbot234]

Scheduling solutions are easy to verify. For other problems, verification would be harder.

[cjbarber]

> Non-technical users expect a CEO's secretary from TV/movies: you do a vague request, the secretary does everything for you. LLMs cannot give you that by their own nature.

What are you using today? In my experience LLMs are already pretty good at this.

> Please for the love of god actually go outside and talk to people outside of the tech bubble.

In the past week I've taught a few non-technical friends, who are well outside the tech bubble, don't live in the SF Bay Area, etc, how to use Cowork. I did this for fun and for curiosity. One takeaway is that people at startups working on these products would benefit from spending more time sitting with and onboarding users - they're very powerful and helpful once people get up and running, but people struggle to get up and running.

> People don't want "personalized interfaces that change every second based on the whims of an unknowable black box". They have plenty of that already.

I obviously agree with this, I think where our view differs is I expect that models will be able to get good at making custom interfaces, and then help the user personalize it to their tasks. I agree that users don't want something that changes all the time. But they do want something that fits them and fits their task. Artifacts on Claude and Canvas on ChatGPT are early versions of this.

[troupo]

> What are you using today? In my experience LLMs are already pretty good at this.

LLMS are good at "find me a two week vacation two months from now"?

Or at "do my taxes"?

> how to use Cowork.

Yes, and I taught my mom how to use Apple Books, and have to re-teach her every time Apple breaks the interface.

Ask your non-tech friends what they do with and how they feel about Cowork in a few weeks.

> I think where our view differs is I expect that models will be able to get good at making custom interfaces, and then help the user personalize it to their tasks.

How many users you see personalizing anything to their task? Why would they want every app to be personalized? There's insane value in consistency across apps and interfaces. How will apps personalize their UIs to every user? By collecting even more copious amounts of user data?

[roel_v]

"LLMS are good at "find me a two week vacation two months from now"?"

Of course they are. I gave one a similar prompt a few weeks ago, albeit quite a bit more verbose (actually I just dictated it, train of thought, with couple of 'eh actually, forget what I just said about x, do y instead") and although I wasn't brave enough to give it my credit card and finalize the bookings, it would have paid for the bookings I had it set up for me, had I done that. I gave it some RL constraints, like "we're meeting friends in place xyz at such and such date, make sure we're there then" and it did everything from watching we wouldn't be spending too many hours driving per day to check that hotels are kid friendly to things to do and see and what public holidays there are so that we know when supermarkets close early and a bunch of details I wouldn't have thought of. It checked my (and my wife's) calendar, checked what I had going on work wise, etc.

That is a fully solved 'problem' man. LLMs will run the whole thing for you. Just provide it with the login details to booking websites and you're off to the races.

I did have it upgrade the car, even if that pushed the cost outside the budget I gave it. Next time it'll know LOL.

[suddenlybananas]

>although I wasn't brave enough to give it my credit card and finalize the bookings

So it's not trustworthy enough for you, someone clearly interested in the hype of LLMs.

[roel_v]

It's a matter of getting used to things. We're only a few weeks further, I maybe would have given it now. It'd need some way to keep it private I guess, maybe I could have used a one off CC number. Those are just technicalities at this point. It got me to the point where I just had to enter my details and click a few confirm buttons. Those are solved problems. I'm not sure why the denialists here are saying those things are 'impossible'. I mean I've seen them happen, what do you want me to say? Claiming this is 'just hype' is ostrich behavior. I've been playing with an abliterated Gemma 4 yesterday on my local machine. Yes it would take longer and require a bunch of harness fiddling, but even if OpenAI and Anthropic would collapse tomorrow, I'm confident I could still do the exact same thing the day after with with what I have right now on my hard disk. I'm not sure what you want me to tell you mate. Yes there's rough edges to work out or just in general workflows to improve but the ideas are way beyond 'proof of concept'. There's people like myself using these things for purposes that 6 months ago were science fiction. I don't care if you believe me or not, I'm just some dude on the internet, but level of delusion on how 'inferior' these models (with proper harnessing) are is mind boggling for someone like me who sees it happen literally 20 centimeters to the side on my screen from where I see people claim that those things are impossible.

[baq]

> Or at "do my taxes"?

codex did my taxes this year (well it actually implemented a normalization pipeline and a tax computing engine which then did the taxes, but close enough)

[William_BB]

> well it actually implemented a normalization pipeline and a tax computing engine which then did the taxes, but close enough

You can't seriously believe laymen will try to implement their own tax calculators.

[baq]

of course not.

what I believe is that laymen will put all their tax docs into codex and tell it to 'do their taxes' and the tool will decide to implement the calculator, do the taxes and present only the final numbers. the layman won't even know there was a calculator implemented.

[tsimionescu]

If your prompt was more complex than "do my taxes", then this is irrelevant.

[baq]

it was many hours of working with codex, guidance and comparing to known-good outputs from previous years, but a sufficiently smart model would be able to just do it without any steering; it'd still take hours, but my input wouldn't be necessary. a harness for getting this done probably exists today, gastown perhaps or something that the frontier labs are sitting on.

[ravenstine]

Sounds fascinating! If you wrote an article on this I bet it'd have a good shot at making it to the home page of HN.

[jeffgreco]

> LLMS are good at "find me a two week vacation two months from now"?

Yes?

===

edit: Just tested it with that exact prompt on Claude. It asked me who I was traveling with, what type of trip and budget (with multiple choice buttons) and gave me a detailed itinerary with links to buy the flights ( https://www.kayak.com/flights/ORD-LIS/2026-06-13/OPO-ORD/202... )

[troupo]

I'd love to try and replicate, but I'm not letting any of these tools anywhere near a real browser and capabilites :)

[mazurnification]

Perfect - and this use case will be enshitificated first. LLM provider will charge small fee for proper recommendation placing. Got to recoup investment.

[a1j9o94]

This is effectively how I treat my AI agents. A lot of the reason this doesn't work well for people today is due to context/memory/harness management that makes it too complex for someone to set up if they don't want a full time second job or just like to tinker.

If you productize that it will be an experience a lot of people like.

And on the UI piece, I think most people will just interact through text and voice interfaces. Wherever they already spend time like sms, what's app, etc.

[trvz]

Most knowledge workers aren't willing to put in the effort so they're getting their work done efficiently.

[louiereederson]

Maybe but the product category is not necessarily a monolith in the same way that Claude Code is. These general purpose tools will have to action across a heterogeneous set of enterprise systems/tools. A runtime environment must be developed to do that but where that of the agent ends and that of the enterprise systems begins is a totally open question.

[bob1029]

> A runtime environment must be developed to do that but where that of the agent ends and that of the enterprise systems begins is a totally open question.

I think something like SQL w/ row-level security might be the answer to the problem. You often want to constrain how the model can touch the data based upon current tool use or conversation context. Not just globally. If an agent provides a tenant id as a required parameter to a tool call, we can include this in that specific sql session and the server will guarantee all rules are followed accordingly. This works for pretty much anything. Not just tenant ids.

SQL can work as a bidirectional interface while also enforcing complex connection level policies. I would go out of band on a few things like CRUD around raw files on disk, but these are still synchronized with the sql store and constrained by what it will allow.

The safety of this is difficult to argue with compared to raw shell access. The hard part is normalizing the data and setting up adapters to load & extract as needed.

[cjbarber]

> Maybe but the product category is not necessarily a monolith in the same way that Claude Code is. These general purpose tools will have to action across a heterogeneous set of enterprise systems/tools.

What would make it not be a monolith? To me it seems like there'll be a big advantage (e.g. in distribution, user understanding) for most people to be using the same product / similar interface. And then the agent and the developer of that interface figure out all the integrations under that, invisible to the user.

[louiereederson]

I mean there is a runtime layer that needs to be developed, and some of it may live in CC/Codex and some might live in the various enterprise systems. Someworkflow automations and some amount of the semantic layer may for instance exist in your CRM/ERP/data platform. Yes the front-end would be owned by the chat interface, but part of the solution may exist in the various enterprise systems. This would be closer to a distributed system than a monolith. The demos and marketing language point to this as the direction of travel (i.e. the reference to Atlassian Rovo, etc.).

[cjbarber]

Thanks for answering!

[eldenring]

I think the coding market will be much larger. Knowledge work is kind of like the leaf nodes of the economy where software is the branches. That's to say, making software easier and cheaper to write will cause more and more complexity and work to move into the Software domain from the "real world" which is much messier and complicated.

[cjbarber]

Yes, and the same thing will happen in non-coding knowledge work too. Making knowledge work cheaper will cause complexity to increase, more knowledge work.

[eldenring]

I don't think so, the whole point of writing software is it is a great sink for complexity. Encoding a process or mechanism in a program makes it work (as defined) for ever perfectly.

An example here is in engineering. Building a simulator for some process makes computing it much safer and consistent vs. having people redo the calculations themselves, even with AI assistance.

[cjbarber]

The history of both knowledge work and software engineering seems to be increasing in both volume and complexity, feels reasonable to me to bet on both of those trendlines increasing?

[visarga]

Yes, I have a theory - that higher efficiency becomes structural necessity. We just can't revert to earlier inefficient ways. Like mitochondria merging with the primitive cell - now they can't be apart.

[joshysmith]

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

[frez1]

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.

[jorblumesea]

really struggling to understand where this is coming from, agents haven't really improved much over using the existing models. anything an agent can do, is mostly the model itself. maybe the technology itself isn't mature yet.

[cjbarber]

My view is different. Agent products have access to tools and to write and run code. This makes them much more useful than raw models.

[visarga]

Yes, I think they unlock a whole new level of capability when they have a r/w file system (memory), code execution and the web.

[flir]

That's not the model, that's the box the model came in.

It's unlikely we've hit the limits on improving agent UX, but there are some fundamental limits on LLMs that seem unlikely to be fixed by better UX.

[croes]

You know what happens to a predator who makes its prey go extinct?

AI is doing the same

[andoando]

Totally agree, AI interfaces will become the norm.

Even all the websites, desktop/mobile apps will become obsolete.

[donnisnoni]

AI won't kill apps, it will just change who 'clicks' the buttons. Even the most powerful AI needs a source of truth and a structured environment to pull data from. A world without websites is a world where AI has nothing to read and nowhere to execute. We aren’t deleting the UI. We’re just building the backends that feed the agents.