All podcasts / Lenny's Podcast / Summary

AI predictions: Job markets, Codex beats Claude, and the death of org charts | Dan Shipper

2026-05-24 - 93 min - source - Read full transcript

Lenny Rachitsky (host)Dan Shipper

Key insights

Work is bifurcating into two surfaces: a company-wide super-agent reached via Slack, and a personal work environment like Codex or Claude Cowork that becomes the operating system for all knowledge work.

Dan predicts most professional tasks (email, documents, research) will happen inside an agent environment with an embedded browser rather than a traditional SaaS UI, with the agent watching and acting alongside the human in real time.

ai-agent-work-surfaces

Personal agents lost out to one central company super-agent because every useful agent needs a human who actively cares for it.

Every initially bet on everyone having their own OpenClaw-style agent, but found the maintenance burden (context rot, breakage, SSH fiddling) too high for most employees; companies like Shopify and Ramp instead run one well-tended super-agent, often maintained by a forward-deployed-engineer-style role, with specialization trickling down over time.

ai-agent-work-surfaces

The SaaS apocalypse narrative is wrong; agents actually increase SaaS usage and buying SaaS stocks is a good bet.

When users run SaaS apps through their own agent's tokens (via CLI or browser-in-agent) rather than the vendor paying for AI, it lowers the vendor's margin pressure while agents make far more requests than humans ever did, spiking demand. Every's own SaaS spend has grown even as the team is AI-pilled.

saas-and-agent-economics

Building for agents changes what software needs: approval queues, action logs, rollback, and interfaces that make sense when an agent can make a billion requests in seconds rather than one human clicking around.

Dan argues the old model of software built purely for a human, or a bare CLI built purely for an agent, is being replaced by products where a human and an agent collaborate on the same piece of work simultaneously, each needing visibility into what the other is doing.

ai-agent-work-surfaces

CLI-first coding agents were a temporary phase, not the endpoint; GUIs are coming back because they are simply nicer to work in.

Dan says Every 'speed ran the CLI era' - most non-programmers at the company have already moved off the command line into GUI tools like Codex, Cowork, and Cursor, and even most technical staff use CLIs only occasionally now.

coding-agent-competition

OpenAI's Codex has taken the lead over Anthropic's Claude Code/Cowork in the coding-agent race, in Dan's daily use, largely because of its in-app browser and general knowledge-work orientation.

Dan calls Codex his daily driver, says OpenAI has 'gotten back the mandate of heaven' after a rough stretch, and singles out the ability to browse and act inside a document (e.g. his markdown editor Proof) as the single most powerful and underrated pattern right now.

coding-agent-competition

The AI job apocalypse is not happening; models make yesterday's human competence cheap and commoditized, which pushes humans to keep finding the next frontier rather than eliminating work.

New model releases rapidly spread a capability to everyone (e.g. anyone can now build a landing page), making that capability valueless because it's commoditized; humans respond by using that frozen competence as a floor to build something new, which is why demand for engineers has risen even as coding got automated.

future-of-jobs-and-roles

Product managers and full-stack designers are the biggest career winners of the current AI wave.

Dan cites Every's own team (a formerly non-technical PM who now ships product using Codex/Cursor, and designers who now submit pull requests directly instead of handing designs to engineers) as evidence that people with strong product or design sense plus light technical fluency are becoming dramatically more effective and independent.

future-of-jobs-and-roles

A new 'forward deployed engineer' role - someone who babysits and maintains internal agents - is a durable new job created by automation, not eliminated by it.

Every runs consulting engagements built around this role; these people spend much of their time in Slack correcting and guiding an internal agent (Claudia) rather than writing code directly, illustrating that automation created a job rather than only destroying ones.

future-of-jobs-and-roles

AI-generated documents, plans, and emails are becoming normal and even preferred, as long as the human author stands behind every line.

Every ran its entire 2025 quarterly planning process through Notion agents interviewing each team, producing plan documents better than most humans would write manually; Dan distinguishes real AI-assisted work from 'slop' by whether it took less time to produce than to read and whether the sender can defend its content.

riding-the-model-wave

The way to stay relevant through model upgrades is to actively re-test your own workflows against every new model release rather than assuming past limitations still hold.

Dan calls this 'riding the models' and 'turning the rock over again' - his senior-engineer benchmark score jumped 30 points between model generations, and he argues the edge of AI usefulness is wherever a model meets a real task, not in San Francisco where models are built.

riding-the-model-wave

Sales and executive roles are the least changed by AI so far, because sales is inherently in-person and CEOs can still choose to delegate AI adoption.

Dan argues this is a temporary illusion for CEOs specifically: a company only goes as far in AI as its CEO personally engages, and that gap will become obvious and costly soon, even though it currently looks optional at the top.

future-of-jobs-and-roles

Books referenced

The Writing Life - Annie Dillard - Required reading for every new hire at Every; Dan says the last chapter captures the intersection of writing, technology, and time that defines the company.
The Second World War - Winston Churchill - Dan is partway through Churchill's history/memoir and admires the kinship of someone who both did the thing and wrote about it.
The Riddle of Angels - unspecified - A book relating Heisenberg's uncertainty principle, Borges, and Kant that Dan calls mind-blowing and full of overlaps with AI.
The Power Broker - Robert Caro - Dan's ongoing obsession, a history of New York and Robert Moses.

Media referenced

The Dark Wizard - other - Mini-series documentary about free-solo climber and BASE jumper Dean Potter that Dan recently enjoyed.
100 Foot Wave - show - Documentary about big-wave surfers Dan mentions in the same vein as The Dark Wizard, about people pushing extremes.
Rob Burbea talks - other - Dan references a meditation teacher's talk on relating to hard things from a position of spaciousness and strength, which he applies to fear about AI changing his job.

Companies

Every - Dan's company; ~30-person AI-native media and software studio running six internal products, used as the running example throughout the episode.
Anthropic - Built Claude Code then Cowork; discussed as the original pioneer of agent-on-your-computer coding tools, now facing competition from OpenAI's Codex.
OpenAI - Discussed via Codex, which Dan says has overtaken Claude Code as his daily driver and 'the mandate of heaven' in the coding-agent horse race.
Cursor - Praised for its cloud implementation but seen as staying narrowly focused on programmers rather than expanding into general knowledge work.
OpenClaw - Open-source personal-agent harness Every experimented with company-wide; found too fragile and high-maintenance for most non-technical users to keep alive.
Shopify - Cited as an example of a company running one central super-agent (nicknamed River) rather than individual personal agents.
WorkOS - Podcast sponsor; enterprise auth/SSO/SCIM platform.
Vanta - Podcast sponsor; compliance automation platform (SOC2, ISO 27001, HIPAA).

Techniques and frameworks

The reach test - Every's internal heuristic for whether a new AI tool is actually useful: do you organically reach for it first thing in the morning.
The senior engineer benchmark - Dan's homemade benchmark comparing models against two human senior engineers rewriting his vibe-coded app Proof from first principles; models jumped from ~30/100 to 62/100 with GPT-5.5.
Riding the models - Dan's core advice for staying relevant: whenever a new model ships, actively re-test it against your own work rather than assuming it still can't do something.
The allocation economy - Dan's earlier framework (from a past essay) for how humans work with AI: like being a manager who allocates and checks in on work rather than doing it directly.

Summary

Dan Shipper, CEO of the AI-native media and software studio Every, returns to Lenny's Podcast a year after correctly predicting the rise of Claude Code for non-engineering work, this time to lay out a structured set of predictions for how work will change over the next year. He frames the shift around two work surfaces: a company-wide "super-agent" reached through Slack that people delegate tasks to, and a personal environment like Codex or Claude Cowork that becomes the default operating system for individual work, complete with an in-app browser so the agent can watch and act alongside the human on documents, emails, and research. Every experimented with giving everyone their own personal agent via OpenClaw and abandoned it: agents need a human who actively tends them, and most employees couldn't sustain that, so the company (like Shopify and Ramp) settled on one well-maintained central agent with specialization trickling down as models improve.

A recurring thread is that the "SaaS apocalypse" and "AI job apocalypse" narratives are both wrong. Dan argues agents actually increase SaaS usage and spend, since running an app through your own agent's tokens shifts cost away from the vendor while multiplying the volume of requests hitting their infrastructure; he says he would buy SaaS stocks. Similarly, he contends models make "yesterday's human competence" cheap and commoditized rather than eliminating the need for humans: once everyone can build a landing page or write serviceable code, that capability stops being valuable, and humans respond by pushing into whatever is still unsolved, which is why demand for engineers has grown even as raw coding became automated. He illustrates this with his own "senior engineer benchmark," built after his own vibe-coded app broke in production: models jumped from roughly 30/100 to 62/100 against two human rewrites when GPT-5.5 shipped, showing real progress but also that models still default to patching rather than the aggressive first-principles rewrites a senior human engineer chooses.

On careers, Dan is emphatically bullish on product managers and full-stack designers, citing internal examples of a formerly non-technical PM who now ships product independently using coding agents, and designers who submit pull requests directly instead of handing off to engineers. He also identifies a new durable role, the "forward deployed engineer" who babysits and maintains internal agents, as evidence that automation creates jobs as often as it removes them. Sales and executive roles, by contrast, are described as the least changed so far, though Dan warns the appearance that CEOs can opt out of engaging with AI personally is temporary and increasingly costly. Throughout, he frames AI-generated writing (plans, emails, guides) as increasingly acceptable and even preferable, provided the human sender stands behind every line, distinguishing that from "slop" that took less time to produce than to read.

The conversation closes with Dan's core piece of advice: "ride the models" by re-testing your own workflows every time a new model ships rather than assuming old limitations still apply, since the real edge of AI usefulness is wherever a model meets a genuine task, not in the labs that build it. In the lightning round he recommends Annie Dillard's The Writing Life, Churchill's history of World War II, a book on Heisenberg/Borges/Kant called The Riddle of Angels, and Robert Caro's The Power Broker, along with the documentary The Dark Wizard, and names Codex as his most underrated current tool.

Notable Quotes

"What models do in general is they make yesterday's human competence cheap. And so it becomes commoditized. It's not valuable anymore." - Dan Shipper

"Automation is a lie. In the sense that every time you automate something, in order to make sure the automation is working well, you need a human on top of it, like making sure that it's working well." - Dan Shipper

"The kind of software that you make for that is going to be very different... you and the agent are using the app together." - Dan Shipper

"I would buy SaaS stocks right now... the SaaS apocalypse is dumb." - Dan Shipper

"The edge of AI is wherever AI meets like a real human doing something." - Dan Shipper