Every Agent Needs a Human

Dan Shipper, CEO of Every, interviewed by Lenny Rachitsky (Lenny's Podcast)

I post one executive summary daily of an interview I enjoyed and learnt from. I loved this interview that @lennysan did with @danshipper. Great insights around building and running a AI-native company for founders and operators.

Summary: Dan Shipper runs what is probably the most AI-forward company in the wild, has doubled headcount in a year, and thinks the AI job apocalypse is wrong. His read after a year inside the frontier: agents do not work without a human running them, the work surface for knowledge work is collapsing into Codex and Claude Co-work, and the people who win are the ones who keep riding each new model into work they could not do before. SaaS is back. PMs and designers are back. Buy the stocks.

1. Automation Is a Lie. Every agent needs a human who cares about it. The minute you sever that connection, the agent stops being useful. Shipper has doubled Every's headcount in a year while running 6 AI products internally, and the math still works because someone has to watch the agent, add context, and fix it when it breaks. Treat "fully autonomous AI" as a marketing line. Staff for the human who keeps the agent working.

2. The Super Agent Won. Shipper used to believe in personal agents on every laptop. He flipped. The pattern that actually works is one company-wide agent with a forward-deployed engineer who keeps it healthy: Shopify has River, Ramp has theirs, Every has Claudie. Build for the central agent first. Push specialization down only as the models become more independent.

3. Codex Is the OS. The work surface for knowledge work is collapsing into Codex and Claude Co-work. Shipper has been at inbox zero for 10 days because Codex pulls every email, opens an in-app browser to research, and drafts replies while he monologues at it. The agent watches him, he watches the agent, and the document, the browser, and the model are all in one window. Anything you ship for knowledge workers in 2026 needs to make sense inside that one window.

4. CLIs Are Over. Most technical people at Every have moved off the CLI back into GUIs for Codex and Cloud Co-work. GUIs were invented for a reason and the benefits of the CLI era port cleanly into a window. Stop bragging that your product is terminal-native. Ship the GUI.

5. The SaaS Apocalypse Is Dumb. Shipper would buy SaaS stocks today. Every is the most AI-pilled company he knows and Every's SaaS spend is up year over year, because agents multiply the number of users a SaaS product serves. The right move for a SaaS founder is to design for humans and agents working on the same surface at the same time. The customer brings the tokens. You keep the margin.

6. The CEO Test. "Your company is only going to go as far as your CEO goes in AI." Shipper sees this with every senior leadership team he works with: CEOs who delegate AI exploration lose intuition for what their org can ship, and the company plateaus there. The CEO has to keep hands on Codex every week, even when it would be easier to hand off to a chief of staff. If you stopped playing with each new model drop, your team will too.

7. Yesterday's Competence, Commoditized. Models make yesterday's human competence cheap. The moment a new model lands, anyone can make a landing page or write a tweet or ship a feature at the median of yesterday's craft, so that work commoditizes overnight. Humans then take the frozen competence and push past it into something the model has not seen. The job is to stay on the edge of that gap, finding what the model has not yet absorbed.

8. Benchmarks Lie About Autonomy. Shipper built a "senior engineer benchmark" by handing models his own vibe-coded codebase and asking them to rewrite it. GPT 5.5 hit 62, up from 30 the prior generation, and headlines would say senior engineers are about to be replaced. The trick is that the prompt itself, telling the model to rip out the vibe-code and start from first principles, took weeks to craft, and a senior engineer would have known that on their own. The benchmark measures what we already learned to articulate. Humans do the part you cannot score yet.

9. PMs Eat the Future. A PM at Every named Marcus took a year off, got AI-pilled, learned Cursor and Claude Code, and now ships faster than anyone on the team while keeping his product instinct intact. A year ago Every could not have hired him for this role. Today he runs Spiral, Every's writing app, solo. PMs with a feel for users, taste for writing, and the discipline to drive a coding agent are the most valuable hire in tech right now. If your PMs have not flipped, that is the gap to close.

10. Designers Ship Now. Designers at Every are opening pull requests directly, because Claude Code makes the gap between mockup and merged interaction collapse. The interactions they design are the exact thing vibe-coding cannot do, which is why every "AI-built" site looks the same. A full-stack designer with strong taste is now a one-person feature team, and the design job market data has not caught up yet. Hire them while it has not.

11. Strategy Docs by Agent. Every ran its 2025 Q4 planning entirely through Notion agents. Each team talked to the agent about goals, metrics, and pushback against top-down strategy. Shipper got back planning docs better than most humans would have written on their own. The new etiquette: send AI-written docs all you want, but stand behind every line. If you cannot answer a question about something your agent wrote, you do not get to send it.

12. Ride the Models. The single instruction Shipper gives anyone worried about losing their job: ride the models. Every time a new model drops, point it at the thing you actually do, find what it cannot do yet, and come back next release. The frontier of AI is wherever a model meets a real human doing a real job. Every is in Brooklyn precisely because that distance from the lab is an advantage. Find your moment of joy with AI, share it, repeat.