Every CEO Dan Shipper: Stop focusing on the interface. In the future, whoever controls the backend data will have the final say.

If you've recently been bombarded with views like "AI will replace everyone", "SaaS is doomed", and "Everyone will have to work in the command line in the future", then this interview with Dan Shipper might turn many of your preconceived notions upside down.

Dan is the co - founder and CEO of Every. This company has only about 30 people, but from editors to designers, and from operations to sales, almost everyone is integrating AI into their daily work. As a result, they seem to be living in the future work style ahead of time.

One year ago, Dan said something in this program that few people cared about at that time: The real value of Claude Code is not to help programmers write code, but to enable non - technical people to use it to process files, organize hard drives, and complete various miscellaneous tasks. One year later, Anthropic launched Cowork, and OpenAI's Codex is also catching up in this direction.

Dan was right in his judgment.

This time, he brings another set of counterintuitive judgments:

SaaS will not only survive but may even become more profitable; the command line is not the future work interface and is actually nearing the end;

When the mainstream discussion is all about automation and interface optimization, he says that the real power center lies not in the operating interface but in the backend system and data itself;

AI will not bring about an employment apocalypse. The new and increasingly important position is called the "forward deployed engineer".

For ordinary people to avoid being eliminated, the most important thing is to learn to "ride the models forward".

Original article link: https://www.youtube.com/watch?v=4D3hDmGhFhA

The following is a compilation.

1

He said a year ago that people underestimated Claude Code

Host: Last time you were on the program, you made a judgment that seemed rather offhand at the time: People underestimated Claude Code, especially its value in non - engineering work, such as organizing files, managing hard drives, and handling various non - technical tasks. That was a year ago, and hardly anyone was saying that at the time. It turns out you were so right. Anthropic launched Cowork, and OpenAI's Codex also started moving in this direction. People gradually realized that these "coding agents" are not just for programmers. Later, you even wrote a newsletter to further explore this idea, asking people "How else can Claude Code be used in non - engineering work?" As a result, you received a lot of case studies, and that article became one of your most popular pieces. So, in today's program, I'd like to continue along this line of thought: What other changes do you see that others haven't noticed but you think will happen soon? Your company, Every, seems like a "company living in the future ahead of time" — so perhaps it's best to start by asking you: How do you and your team actually work?

Dan: Thank you for the introduction. Actually, I've always thought that the least interesting way to "predict the future" is to make baseless prophecies. A more effective method is not to "guess" but to experience it first. Every currently has about 30 people. When you interviewed me last time, we had about 15 people, so the number has doubled in a year. Our team includes engineers, designers, writers, editors, salespeople, and customer service representatives. Everyone is an early adopter of AI. There's a strong common trait among us: a willingness to try new things, a curiosity to experiment, and a genuine desire to integrate AI into our work. The result is that a small "future showcase" has formed within our company. In most companies, there are usually different types of people: those who adopt new tools early, those who catch up slowly, and those who are clearly resistant. But here, almost everyone is actively experimenting, so many changes become apparent earlier. Another crucial point is that since we are constantly evaluating models, writing about AI, and discussing AI, we get access to beta and alpha versions earlier than many people, allowing us to experience things that model companies haven't officially released yet. In a way, we're not just observing trends but also participating in the process of their formation.

So, for me, "seeing the future" is actually about two things: First, you're actually using it on the front line; second, you can clearly articulate and write about the changes you see. Many times, when you clearly express a vague change, it becomes more real, not only for you but also for your team and external readers. That's how the thing about Claude Code came about. As soon as it came out, we tried it. That's what we do — we try new things as soon as they arrive. At first, it was still in its early stages, but around Sonnet 3.5 or 3.7, when we did an "intuitive test" internally, we suddenly felt that something had changed. It had reached a turning point in terms of its capabilities. What shocked me the most was that it almost removed the "code editor" layer. From that point on, we gradually shifted towards a new way of collaboration within the company: We currently have about six software products running internally, while there were only two or three at that time. Many people in the team don't look at code much anymore. Instead, they directly talk to their computers in natural language and let Claude Code do the work for them in the terminal. Then I started to realize that things had changed. As someone who likes to "take one more step forward and see", I naturally started asking questions: Can it also help me with writing? Can it do research? Can it handle other knowledge - based tasks? You'll find that many scenarios aren't fully developed yet but are already "good enough". We have an internal term called "reach test": When you wake up in the morning, do you instinctively reach for it? If a tool passes this test, it's no longer just "interesting" but is starting to become part of your work environment.

Host: Your combination is really powerful: On the one hand, your whole team uses the latest things; on the other hand, you're very good at identifying "what's new, strange, important, and that others haven't realized yet". Coupled with your ability to write and express, this combination really allows you to see changes earlier than others.

2

Every company will have an AI agent for repeated conversations in the future

Host: Okay, let's get to the main point. You said you want to divide today's judgments into three categories. Let's start with the first one: How will the way we work change in the next one or two years? I really like this question because if you only look at benchmarks, it's easy to come to a scary conclusion: AI will become increasingly capable of autonomously completing long - term tasks. Will our jobs soon be taken away? For example, benchmarks like METR measure "how long a new model can autonomously complete tasks". When you see such data, it's quite astonishing: The latest models may already be able to complete tasks at the 17 - hour level with a 50% accuracy rate. It sounds really scary. And to be honest, this trend is real, and the capabilities of models are indeed rising rapidly. But my real experience and judgment are that when we look back in a year, we'll find that there isn't less work for humans to do; instead, there's more. There's a very interesting paradox here.

Dan: So, if you ask me: What's the biggest change in the way we work in the future? My answer is: Work will split into two paths. The first path is that in every company, almost everyone will have frequent conversations with at least one AI agent and delegate a part of their work to it. The second path is that the main interface for you to actually complete your work will gradually shift to environments like Codex, Claude Code, and Cowork. That is, on one side, there's the "company - level agent" within the organization; on the other side, there's your "personal workbench" on your own computer. Let's start with the first one. When people used to imagine "AI entering the company", they often first thought that everyone would have their own agent. But now, I think what's more likely to happen in the short term is that each company will first have a "super - agent", a super agent shared by the whole company. You can already see some companies starting to do this. For example, Shopify famously did one, and Ramp also has one now. Behind this question, there's actually a divergence at the architectural level: Should each person have an agent? Each team? Or the whole company? Will agents develop like a "parallel organizational structure"? Will it become some kind of "shadow org chart"? When OpenClaw first came out, I really believed in the idea of "one agent per person". That idea was very appealing: Everyone would have an agent by their side, like a clone or a little elf on your shoulder, knowing your habits and acting on your behalf. If you've seen The Golden Compass, it's like everyone has their own dæmon, a part of their soul following them. I really believed in that direction back then. But later, I completely changed my mind.

Now, I'm more convinced that at least at this stage, a more realistic model is "one super agent per company". Why? Because you'll soon find that setting up a personal agent is really troublesome. Whether it's OpenClaw or other harnesses, people will get excited at first: "This is so cool. I'll set one up too." But then they'll quickly face reality: You need to deploy it, set up SSH, handle permissions, monitor its running status, and it often breaks down. Once it breaks, no one wants to fix it. Most people don't want to spend their energy on "maintaining this agent", or they simply don't have the time and willingness. And for the current agents, there's an underlying reality: They need someone to constantly care about them. It's not that once they're set up, they'll grow automatically. For them to be truly useful now, often someone needs to have a "personal relationship" with them: constantly monitor what they're doing, help them fix problems, and ensure they can be valuable to the business. Once this relationship is broken, the agent will quickly degenerate into something that "can be used occasionally but is generally unreliable". So, at this stage, the most feasible model is not "everyone has an agent" but rather: First, a person of the forward deployed engineer type is responsible for maintaining the company - level agent. This kind of person will ensure that it's truly available to the whole company. Then, based on this, team - level agents and personal - level agents will gradually emerge. In the long run, I still believe that personal agents will come. But that will be based on the premise that the models are more stable and independent and don't require you to take care of them every day. So, my current judgment is: Start with an agent at the top of the company and then differentiate downwards. The core mechanism behind this is a simple statement: Agents need people to care about them.

Host: This point is very interesting. You said that agents need "gardening", constant care. I totally resonate with this. Because you need to supplement the context, connect the permissions, and fix the errors. Once it gets too troublesome, people will give up and go back to simpler ways.

Dan: Exactly, that's right.

Host: Okay, then let's talk about the second path. You just said that most future work will take place in environments like Codex, Claude Code, and Cowork. I really want to explore this part.

Dan: I also think this is the most interesting part. Anthropic realized early on that if you put an agent on the user's own computer, it naturally has a very powerful combination of capabilities. Since it runs on your computer, it means it has the same permissions as you, can use the terminal, and has a strong understanding of the terminal because there's a lot of content on the internet about CLI, shell, and toolchains. As a result, a very powerful new paradigm has emerged. At first, people regarded the coding agent as a "pair programming assistant", meaning it would help you fill in a few lines of code beside you. But Anthropic was one of the first companies to truly turn it into an "agent that can do things for you on your machine". Of course, there were earlier attempts, such as Devin's cloud environment approach, and OpenAI also tried a similar direction, but it was "putting the agent on your computer" that really achieved large - scale adoption. Later, they further discovered that once a coding agent is on your computer and can build anything, it's not just a coding tool. It can also become a general knowledge work tool. So, people started using Claude Code in a "hacky" way for various tasks. Anthropic's later development of Cowork was actually putting a more user - friendly shell on this idea for non - engineering users. OpenAI took a different path. In my opinion, the early Codex was too technical: very smart but a bit "literal in understanding your instructions". It would mechanically execute whatever you said.

But around a few months ago, especially when they released version 5.3, I felt that OpenAI also clearly shifted its direction: They started developing an agent more suitable for general knowledge work rather than just a system for programmers. Then they released the desktop version of Codex. I think they actually saw the path from Claude Code to Cowork by Anthropic and then chose to be more radical and take a big leap forward. If you ask me what I'm mainly using now, to be honest: Codex has become my daily driver. I also switch back and forth and use Cowork occasionally, but currently, I think OpenAI has grasped the paradigm of the "workbench" very well. The most crucial point for me is that most of the work you do in the future will be completed in this kind of interface. For example, when I'm writing a document, the desktop version of Codex has a built - in browser. I'll open a thread for each project and then directly open the document page in the application. I usually write in our own online markdown editor, Proof, and let Codex "watch me work" on the side. It can see what I'm doing, and I can see what it's doing. It's not a tool waiting for you to send a prompt in another window but a real collaborative partner working side by side.

This experience is very much like having a work partner who is always online, can take action, and can conduct research. It can respond to you, help you write, help you search, help you operate the browser, and help you mobilize the resources on your computer. I do almost everything this way now. For example, I've kept my inbox at zero for the past 10 days — if you know me, you know this is almost impossible for me. The reason I can do this is that I let Codex cooperate with our email agent, Cora: First, it aggregates all my emails and then renders them into a page. I dictate to the page: "Research this" "This is a lawyer's question. Collect all the documents from the past four years, organize them into a report, and send it out" … and it just does it. There were many things I used to procrastinate on, but now I don't procrastinate as much. Because you're no longer facing a bunch of tasks alone; there's an execution body beside you that can immediately share the burden. So, I used to think that the optimal AI experience would be to put AI into the browser. But now, I'm increasingly thinking that the truly effective way in the future is the opposite: Put the browser into AI. Let AI become the work environment itself and let it see your entire work scene. This isn't common yet. For example, Claude Code itself still can't browse external web pages internally. But I think it will become a very common way of working within a year.

Host: This is actually very profound. You mean that instead of adding an AI button to SaaS, in the future, users may open SaaS in Codex or Claude Code, and the AI workbench will become the main entrance.

Dan: Yes, and there's a big second - order effect behind this. For example, when I'm working in Proof, PostHog, or any other website, if I open it through an environment like Codex, the agent uses my token, not the SaaS vendor's token. It's not the application provider covering the model cost for me. This will, in turn, rewrite the product design logic of SaaS. What you need to do is no longer "necessarily integrate AI into the product" but rather: Enable the agent to access your product smoothly, make the HTML structure friendly to the agent, ensure that changes at the CLI layer are immediately reflected in the UI, and keep the operations of human users and the agent in sync. After doing this, you may not necessarily need to build your own "AI front end". Take Proof as an example. When users bring their own AI to use it, we don't have to