Lore

AI agents

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

Dan Shipper (CEO of Every) argues that AI will not cause a mass job apocalypse; instead, work is bifurcating into (1) everyone having at least one agent they talk to and delegate to, and (2) most professional work happening inside an agentic "operating system" like Codex or Claude Cowork, with humans still required to maintain every agent — and PMs plus full-stack designers positioned as the biggest near-term winners.

Lenny's Podcast · 2026-05-24 · English

Key ideas

  1. Central thesis: AI won't cause mass job loss because automation increases the need for human oversight and demand for skilled workers rises rather than falls, evidenced by Every's own headcount doubling despite being maximally AI-forward.

  2. Predicts work bifurcates into two modes — everyone having at least one agent they delegate to, and most work happening inside an agentic environment like Codex or Claude Cowork that becomes the operating system for all work — framed as the clear emerging competitive landscape.

  3. SaaS won't disappear; agents expand SaaS's addressable market by multiplying effective users, but vendors must restructure their product surfaces (agent-friendly CLIs, browser-embeddable UIs) to remain the interface layer as agents become the primary work surface.

  4. The rapid rise and equally rapid obsolescence of CLI-centric work (Claude Code's early success giving way to GUI-wrapped Cowork and the Codex desktop app within about a year) shows the 'AI-native interface' is a moving target, not a fixed endpoint.

  5. Rising AI capability on benchmarks doesn't map cleanly onto job replacement, because benchmarks only score pre-framed problems, while much of real expert work is the act of framing/deciding what needs to be done in the first place — an unmeasured, undelegated human function.

  6. As non-technical staff gain the ability to produce technical output, the organizational bottleneck shifts from production capacity to integration/curation capacity, which paradoxically increases demand for engineers and data specialists as reviewers and system-builders.

  7. PMs and full-stack designers are argued to be the biggest winners because AI collapses the gap between having taste/judgment and being able to execute directly, letting them ship without needing to marshal a full team.

  8. Resistance to AI-generated writing is argued to be misplaced in most internal contexts, because the real quality bar is whether the author understands and stands behind the content, not who typed it.

  9. Because frontier model access is broadly and cheaply available rather than confined to Silicon Valley or big companies, competitive advantage comes from sustained, curious, hands-on usage rather than proximity or capital.

  10. Closing meta-argument: predictions about the AI future are compared to medieval intuitions about what lies past the horizon, so the source frames its own claims as provisional and explicitly invites being scored against reality a year later.

  11. The reach test — Every's internal heuristic for judging whether a tool is genuinely useful — whether you organically reach for it first thing when you wake up. Apply: Before committing to a new AI tool, notice whether you instinctively open it as your first action of the day; if not, it likely isn't earning real adoption.

  12. One-agent-per-company ("super agent") model — An architecture where a company runs a single shared, actively-maintained agent (e.g., Shopify's 'River') rather than giving every employee a personal agent, because personal agents require too much individual upkeep to stay reliable. Apply: Roll out one well-maintained shared company agent plus a small dedicated team to keep it working, rather than trying to give every employee their own agent.

  13. Parallel shadow org chart — Dan's concept that as work and personal agents proliferate, they form a structure that runs in parallel to and mirrors the human organizational chart. Apply: When mapping out agent deployments across a company, chart them against the existing org chart to spot where agent responsibilities duplicate or leave gaps relative to human roles.

  14. Forward-deployed engineer role — An emerging job function responsible for keeping a company's shared agent functional and useful, grounded in the claim that 'every agent needs a human' who cares about it. Apply: Assign a specific person or small team ownership of maintaining and tuning the company's core agent rather than leaving it unowned.

  15. Computer errands — Every COO Brandon Gall's term for personal, non-work tasks (like ordering groceries) that people delegate to a personal AI agent, kept distinct from work-agent use. Apply: Keep personal-agent use cases (errands, life admin) separate from work-agent use cases when designing or rolling out agent tools, since users prefer that separation.

  16. In-app browser embedded in a coding agent — Putting a web browser inside the AI agent you already use (e.g., Codex) so it can watch and collaborate on websites/SaaS tools live, reversing the earlier idea of putting AI inside a browser. Apply: Use an agent's built-in browser to have it observe and act on SaaS tools directly inside your existing agent thread rather than switching context to a separate app.

  17. Agent-to-agent communication ("two agents are better than one") — Letting one agent talk to another agent to transfer richer user context automatically than a human could type manually. Apply: Design onboarding/support so that when a user arrives via their own agent, your product's agent pulls context directly from that agent instead of asking the user to fill out forms.

  18. Agent-friendly SaaS design principles — A recommended design pattern for SaaS survival: provide an agent-friendly CLI, parseable HTML, and real-time UI reflection of CLI-driven actions so outside agents can operate the product using the user's own tokens. Apply: Build a CLI/API surface and make the UI mirror CLI actions live so a user's personal agent can operate the product directly on the user's own account/tokens.

  19. New UX primitives for agent-driven software — A set of interface affordances needed once agents can take large numbers of actions at once: approval flows, a summarizing inbox of pending/completed actions, logs, and fast rollback. Apply: Add an approvals queue, an action digest, and one-click rollback to products meant to be used by both a human and their agent simultaneously.

  20. Agent-filed bug reports — Letting a user's agent file bug reports directly as structured issues (e.g., GitHub issues) with exact repro steps and code-level diagnosis, instead of the user emailing support. Apply: Enable users' agents to open structured issues directly in your tracker, then dispatch an agent to attempt the fix, closing the loop without human support intervention.

  21. Senior engineer benchmark — Dan's custom benchmark comparing a model's ability to rewrite a 'vibe coded' codebase from first principles against two independent human senior-engineer rewrites, scored out of 100. Apply: To test whether a model has senior-engineer judgment rather than just patching symptoms, have it rewrite a real messy codebase from first principles and score the output against a human expert's rewrite.

  22. Riding the models — Dan's closing motto: continuously adopt each new model release for whatever work you do, rather than avoiding AI out of fear of obsolescence. Apply: Whenever a new model ships, deliberately try it on your actual workflows to discover what new capabilities it unlocks for your specific situation.

  23. Turning the rock over again — The practice of periodically retesting a task a model previously failed at, since capability gaps close as new models release. Apply: Keep a personal list of tasks AI has failed at before and retest them against each new model release to catch newly-crossed capability thresholds.

  24. Edge of AI reframing — Dan's claim that genuine AI innovation happens wherever AI meets a real human doing something specific, not necessarily in SF labs, since lab builders don't fully know how to use their own models. Apply: Don't assume proximity to AI labs confers advantage; look for use cases in your own domain where you can be first to apply a new model capability.

  25. Allocation economy — A concept from an earlier Dan Shipper essay describing humans working with AI the way a manager works with employees — allocating and overseeing rather than doing the work directly. Apply: Frame your role with AI tools as a manager allocating tasks and checking work quality, not as someone replaced or made idle.

  26. Slop vs. good AI-generated content heuristic — A distinction where 'slop' is content that took less time to create than to read and that the author doesn't stand behind line-by-line, while good AI-generated content is fine as long as the author understands and can defend it. Apply: Before sending or publishing AI-drafted material, check whether you could defend every line if questioned; if not, treat it as unfinished.

  27. Org-wide data science bot with permission-awareness — A pattern observed at an unnamed large company: a data-science chatbot wired into the data warehouse, aware of per-user data permissions, maintained by a dedicated team to absorb routine questions. Apply: Build a permission-aware Q&A agent over your data warehouse staffed by a maintaining team, freeing domain experts from fielding basic repeated questions.

  28. Notion agents for quarterly planning — Every's practice of running quarterly planning through Notion agents — one top-level 'company strategy' agent plus per-person interview agents that gather goals/metrics and tie them to overall strategy. Apply: Use a top-level strategy agent plus individual interview agents to collect and synthesize planning input automatically, producing draft strategy documents for leadership review.

  29. METR benchmark — An external benchmark cited in the episode that measures how long the newest AI models can autonomously perform tasks (e.g., '17 hours at 50% accuracy'). Apply: Use METR's autonomous-task-duration scores as a reference point for tracking how far models have progressed toward long-horizon independent work.

  30. Position of spaciousness and strength — A framework the host attributes to Rob Brezsny for relating to difficult situations from calm inner strength rather than forcing through them, contrasted with a 'David Goggins'-style push-through approach. Apply: When facing a hard task or period of uncertainty, consciously adopt a calm, spacious stance rather than gritting through it by force.

Insights

Despite going all-in on AI, Every's headcount doubled in the past year — direct evidence the speaker offers against the intuitive assumption that heavy AI adoption shrinks teams.

SaaS vendors letting a user's personal agent operate the product with the user's own AI tokens, rather than the vendor's, is flagged as a shift that could reshape SaaS margins and reduce vendors' need to build their own in-product AI features.

The qualitative leap Dan measures between models on his own benchmark isn't raw ability but willingness/agency to discard and rewrite code rather than patch around it — a behavioral trait, not just a skill trait.

The biggest observed organizational effect of 'everyone can now code' is a spike in pull-request volume that must be triaged at massive scale — one cited example runs ~50,000 Codex instances to process thousands of daily PRs down to about 1,000 merges — meaning production is no longer the bottleneck, curation is.

CEOs and executives are named as a lagging AI-adoption category specifically because company-wide AI adoption is capped by the CEO's own personal hands-on usage — a non-delegable constraint most companies aren't yet acting on.

The claim that the mass-unemployment scenario 'some AI CEOs are talking about... is not going to happen' directly rebuts doom narratives from unnamed competing AI executives, positioning this episode's forecast as a deliberate counter-narrative.

Real-world examples cited (Shopify's agent 'River,' Ramp's unnamed agent) show major companies already converging on single shared company-wide agents rather than per-employee personal agents, contradicting the 'agent for everyone' hype Dan says he believed in a year earlier.

«The AI job apocalypse is not really a thing. I am super super bullish on PMs and full-stack designers.»

— 00:10

«I'm simultaneously extremely AI pilled and very bullish on humans. Automation is a lie. Every agent needs a human.»

— 00:24

«Creativity. It just feels like it's going to be more and more valuable to stand out from all the slop that people are shipping and launching constantly.»

— 00:34

«What models do in general is they make yesterday's human competence cheap. And so, it becomes commoditized. It's not valuable anymore. What humans do is we go in there and we're like, "Yeah, we have all this frozen human competence from yesterday. How do I use this like make something new and interesting?"»

— 00:40

«I think the SaaS apocalypse is dumb. I would buy SaaS stocks right now. What agents do is increase the number of users of SaaS, not get rid of it.»

— 01:17

«We speed ran the CLI era. It was nice while it lasted, but I think CLIs are over.»

— 01:26

«In order for an AI agent to be useful right now, it really needs a human who cares about it.»

— 14:50

«GPT 5.5 is the only model... that has the sense of agency and confidence to just like rip out old code and just like actually rewrite from first principles.»

— 42:42

«Automation was supposed to take away jobs, but it looks like it just created one or many.»

— 54:50

«Suddenly, everyone's an engineer. That doesn't mean we fire the engineers. There's like way more demand for engineers cuz you need the engineers to like figure out, 'Okay, this is all slop. How does this actually... How should this actually go in our code base?'»

— 75:41

«The only thing you need to do is ride the models.»

— 76:29

«I think the edge of AI is wherever AI meets like a real human doing something... Because the people in San Francisco, they're making it, but they don't actually know a lot about how to use it.»

— 78:41

«There's some really cool things, there's some not cool things, and it's just another horizon.»

— 82:29

Reception

Most commenters found the episode insightful and engaging, with praise for the guest's contrarian takes and practical AI/coding insights, though a vocal minority pushed back on the guest's AI-heavy 'vibe coding' approach as reckless or overhyped.

This is a practitioner's first-person forecast built on Dan Shipper's own company's internal AI usage and self-designed benchmarks rather than external data — a grounded, specific read on where agentic AI work is heading, but its predictions rest on n=1 anecdotes and are explicitly framed by the source itself as provisional, to be scored a year later rather than settled fact.

94:06

↳ Lenny's Podcast · YouTube

Watch original