Ride the Models: Dan Shipper's Bets on Super-Agents, Codex, SaaS and the Human Jobs AI Creates

Dan Shipper runs Every, a media and software company that has grown from 15 to almost 30 people in a year. He says the way to predict the future is not to prognosticate but to live in it together. Everyone at Every, from engineers and designers to editors, salespeople and customer service staff, is an AI early adopter who likes to experiment.

That makes the company a pocket of the future: most firms have a mix of enthusiasts, a middle pack and skeptics, while Every is all enthusiasts. Because it reviews models and acts as a tastemaker, it also gets early access to new releases and helps steer them. Shipper adds that writing about what you notice is part of the method, because articulating the future helps make it real.

Shipper's early call on Claude Code for non-technical work came out of this loop. When Every tried the model around Sonnet 3.5 or 3.7, the team realized the code editor was gone, and people started talking to their computers in English. He describes an internal "reach test": do you reach for the tool organically when you wake up? He and host Lenny Rachitsky agree to revisit the predictions in a year.

Shipper's central prediction is that work will bifurcate. Everyone will have at least one agent they can offload work to, and most of their actual work will happen on their own computer in an environment like Codex or Claude Cowork. He acknowledges the benchmarks suggest AI will take all our jobs, citing a METR chart on tasks of 17 hours at 50% accuracy, yet expects that a year from now humans will still have a lot more work to do.

On the first mode, he has changed his mind. When OpenClaw appeared, he believed everyone would have a personal agent, a kind of daemon on the shoulder. Now he thinks the model for the moment is one super-agent for the entire company, as Shopify and Ramp have, because personal agents break constantly and most people don't want to SSH into a server to fix them.

The underlying reason, he says, is that an agent needs a human who cares about it and watches what it does. Once that connection is severed, the agent stops being useful. So the ideal setup is a forward deployed engineer, or someone with that profile, responsible for keeping the company agent working, with specialized team agents trickling down later and personal agents returning as models become more independent. He expects this agent to live mostly in Slack.

The second mode is the desktop agent. Shipper explains that Anthropic found that an agent running on your own computer, with terminal access to everything you can reach, was far more powerful than cloud-hosted attempts, and the community then hacked Claude Code to do all kinds of work, leading to Cowork. In his view OpenAI was far behind and has recently surpassed Anthropic, with Codex now his daily driver, though he calls it a horse race.

He describes working with a Codex thread per project, an in-app browser, and Codex watching him write in Proof, the markdown editor he built. He has been at inbox zero for about ten days by having Codex gather his emails through Cora, Every's email agent, then dictating instructions on each one. Tasks he used to procrastinate on, like pulling four years of documents for the lawyers, now just get done.

His surprise is the direction of the combination: the winning pattern is not AI inside a browser but a browser inside the AI agent you already use on your computer. Because the agent is running on his tokens, SaaS tools are used through it without the vendor paying for inference. Shipper says this is uncommon now, since Claude Code does not browse external sites inside the tool, but he expects it to be common in a year.

Asked about Cursor, Shipper says it has chosen a lane for programmers and that may limit how far it goes into general knowledge work, a decision he calls hard now that it has been essentially acquired by SpaceX. More broadly he says every model company has realized it needs a harness around the model, with Anthropic offering Claude managed agents and OpenAI expected to respond.

For product builders, the shift is from software meant for a human, or for an agent through a CLI, to software where both work on the same piece of work. That simplifies products, since Proof doesn't need page breaks or table tools when the agent does the formatting. But it also needs new affordances: approval inboxes, logs, quick rollback, and infrastructure that survives agents making a billion requests in three seconds, which he says is already stressing GitHub.

He also says CLIs are over as a main work surface. They won't disappear, but he thinks Claude Code's popularity was never because it was a CLI, and Every's technical people mostly use GUIs such as Codex now. Two agents are better than one, he adds: when a user's Codex talks to an app's agent, it can pass along rich context about the person, and a bug report from an agent comes with exact repro steps that can flow into a GitHub issue and be fixed by another agent.

Shipper argues SaaS companies racing to add their own agents are misreading the moment. If the user's agent is the main work surface, the vendor doesn't pay for the tokens, and the task becomes making software that humans and AI both want to collaborate on. He says that is hard but far cheaper than assuming everyone spends the vendor's tokens, so it may actually save their margins.

He goes as far as saying the SaaS apocalypse is done and that he would buy SaaS stocks, with the caveat that it isn't investment advice. His evidence is internal: everyone at Every has agents and uses Codex, yet the company still pays for plenty of SaaS and its spend is rising year over year. Agents, he says, increase the number of users of SaaS, which means a spike in demand and tough infrastructure and pricing challenges.

Every has doubled its headcount, which Lenny notes is not what people expect from an AI-forward company. Shipper's answer is that automation always needs a human on top to check it works. He recalls writing about the allocation economy, where working with AI resembles being a manager, and notes that managers spend a lot of time working. Benchmarks make AI look more autonomous than it is, which explains why he is working more even as he has more automation.

To make the point concrete he built a senior engineer benchmark. His vibe-coded app Proof kept going down after launch, and Codex's fixes caused other errors until he hired two senior engineers to rewrite it. Now he gives each new model the prompt to rewrite the vibe-coded slop from first principles. Models scored about 30 out of 100 until GPT 5.5 reached 62, using an Opus 4.7 plan, while a human senior engineer scores in the high 80s or low 90s.

He says GPT 5.5 is the only model with the agency to rip out old code instead of patching around the edges. But he predicts every coding model will still take an instruction like "fix these issues" literally a year from now, while a senior engineer would look at the codebase and say it needs a rewrite. Benchmarks rise on problems that are framed and scorable, and the act of framing them is human work. When people compare AI to humans, he adds, it is really one human using AI against another.

Shipper says pull requests are skyrocketing as consultants, ops staff and editors start making them. That creates pressure on whoever has to absorb all the new code. He cites OpenClaw's creator getting thousands of pull requests a day and spinning up many Codex instances to sort them. When building gets easy, the question becomes whether something fits the whole and what to delete, and he praises Anthropic for deleting a lot from Claude Code.

Lenny raises the confusion about roles when engineers design and PMs code. Shipper thinks it will settle, with marketers still doing marketing even if they touch the website, though generalists can now go much further. He also sees new roles: the forward deployed engineer, who keeps agents working, is real. Every has several, including Nitesh, who spends most of his time talking to Claudie, the agent that runs the consulting practice.

Lenny adds that a data scientist friend now mostly reviews bad analysis from colleagues. Shipper says the answer is a team that builds an agent for the basic questions, hooked up to the data warehouse with permissions, which lets data scientists focus on deeper work. He prefers framing the forward deployed team's job less as babysitting and more as an interesting engineering challenge: building systems so less technical people can do formerly technical work safely.

Asked which role has changed least, the two land on sales, where AI helps with top-of-funnel sourcing and research but the core work is still personal, and Shipper also suggests CEOs and middle managers can still get by. He disagrees with the idea that CEOs can delegate AI, because he argues a company only goes as far as its CEO goes. As an example, he had Codex find a head of L&D candidate who had worked at General Assembly and followed him on Twitter.

Shipper predicts we will read much more AI-generated writing in documents and emails. Every ran its quarterly planning with Notion agents that interviewed each employee and produced plans he found incredibly good. His rule is that slop is a document that took less time to make than it takes to read and that the sender doesn't stand behind; an AI-drafted document is fine if you can discuss every line.

Most of his email is now drafted by GPT 5.5 in Codex, and one sent without review turned out to be exactly what he would have written. He still cares about human writing, and Every labels AI contributions and publishes guides meant for both humans and agents. Agents can read thousands of pages instantly, so humans get the story and core ideas while the agent holds the details to apply later.

Shipper is, in his words, super bullish on PMs. His evidence is Marcus, a PM by training who previously ran Axios's writing product and now runs Spiral, Every's writing app. He is lightly technical and got deeply into Cursor during a year off, and Shipper says Every could never have hired him for this job a year earlier. Combining some technical knowledge with spiky product sense and a feel for users, he ships faster than almost anyone and feels liberated from organizing a team.

Lenny agrees: the building is done for you, and what matters is figuring out what to build and what problems to solve. Shipper's second group is full-stack designers. Designers who live in these tools can now build the interactions they imagine instead of handing them off, and since most AI-made UI looks the same, distinctive design stands out. Lenny cautions that designer job postings haven't grown, so it is worth watching whether that changes.

Shipper doesn't buy mass unemployment. He says some reorganizations blamed on AI are really over-hiring, and the pattern he sees is that models make yesterday's human competence cheap. Everyone can suddenly make a landing page or a slop tweet, and since everyone uses the same models the output looks alike. It gets commoditized, and humans then use that frozen competence to make something new for their particular situation.

He argues that, structurally, the models will always trail the people pushing ahead of them, partly because of the incentives to keep models compliant and aligned. Engineers show the pattern: when everyone is an engineer there is more demand for engineers to decide how the slop fits the codebase. What you need to do to stay employed, he says, is ride the models: use them for whatever you do, and try each new one rather than ignore it out of fear.

In practice that means being curious and playful and "turning the rock over" again whenever a model ships, as he did with the senior engineer benchmark. He thinks the edge of AI is wherever it meets a real person doing something, not San Francisco, and says Every in Brooklyn is far ahead of many Bay Area teams because it uses AI for everything. He finds it remarkable that anyone with some money can access the best models the day they appear.

Lenny sums up the paradox: SaaS continues, jobs aren't disappearing and we still use Slack and email, yet every role has been transformed. Shipper compares our intuitions about the future to medieval fears of dragons over the horizon: you arrive and find some cool things, some not cool things, and just another horizon. He thinks it's important not to let your mind tell a story that sounds real now but is more complicated later.

He adds that AI companies were good at scaring people about dangers and he thinks that strategy is ineffective and also wrong, even if the concern is genuine. His closing advice is to try all your workflows in Codex or Cowork, get comfortable with both an async agent and a desktop one, and do it for fun rather than FOMO. Citing a guest from Lenny's show, Nikhil Singhal, he says to find your moment of joy with AI, then keep building.