Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
The AI paradox: More automation, more humans, more work | Dan Shipper
TL;DR
Dan Shipper thinks the 'AI job apocalypse' is mostly wrong — Even as models improve fast, Every doubled from about 15 to nearly 30 people in a year because 'every agent needs a human' to maintain, guide, and integrate the work.
Work will split into two main AI surfaces — Shipper predicts most companies will use one shared 'super agent' in Slack for async delegation, plus a desktop environment like Codex or Claude Co-work where most day-to-day knowledge work actually happens.
The CLI boom was a phase, not the destination — He says 'we speed ran the CLI era' and expects GUIs like Codex desktop to win because humans and agents need to work side by side with shared visibility, approvals, logs, and rollback.
He's aggressively bullish on SaaS, not SaaS extinction — Shipper's contrarian take is that agents will increase SaaS usage, not replace it, because users will bring their own tokens through tools like Codex while apps adapt to be used by both humans and agents.
The winners are PMs and full-stack designers — His strongest talent prediction is that product managers with strong taste and AI fluency, plus designers who can ship their own pull requests, become unusually powerful because they can turn judgment directly into product.
Benchmarks overstate autonomy because they miss the human framing layer — Shipper's 'senior engineer benchmark' showed most models scored around 30/100, GPT-5.5 jumped to 62/100, but he argues the real missing piece is the human who recognizes when the whole codebase is 'slop' and needs a rewrite.
The Breakdown
Dan's edge: don't predict the future, live inside it
Shipper opens with the idea that Every doesn't really 'forecast' AI so much as inhabit it early. The company has grown from roughly 15 to nearly 30 people while making everyone — writers, editors, sales, support, designers, engineers — experiment aggressively with new tools, which gives them a little 'pocket of the future' to observe from.
The first big prediction: one company super-agent beats personal agents, for now
He says most companies will land on one shared agent in Slack before they get a whole shadow org chart of personal agents. He originally believed in one-agent-per-person — like a Golden Compass-style daemon on your shoulder — but flipped after seeing tools like OpenClaw prove too fiddly for most people to maintain. His core line is the memorable one: every agent needs a human who cares about it.
The second big prediction: your real workspace becomes Codex or Claude Co-work
Shipper thinks the more radical shift is that most work moves into a desktop AI surface where the agent can see your browser, documents, files, and terminal all at once. He describes using Codex as a 'parallel work buddy' for writing in Proof, triaging email, researching legal questions, and staying in inbox zero for 10 days straight — which Lenny rightly treats as evidence of something seismic.
Why this changes SaaS instead of killing it
Lenny pushes on the obvious implication: if work happens inside Codex, then SaaS tools run inside the agent, not the other way around. Shipper's response is classic contrarian Dan — 'the SaaS apocalypse is dumb' and he'd buy SaaS stocks — because agents don't eliminate software, they create more users of it, including agent-users. The new challenge for software companies is building products humans and agents can use together, with approvals, logs, rollbacks, concurrency, and agent-friendly interfaces.
Why more automation can still mean more people and more work
This is the heart of the paradox in the title. Shipper says benchmarks make AI look more autonomous than it feels in practice, and tells the story of vibe-coding his app Proof until it kept crashing after launch, leaving him sleep-deprived with 'vibe coder elbow' and bursitis. His homemade benchmark compared models against senior engineers rewriting the mess; most models scored around 30/100, GPT-5.5 jumped to 62/100, but humans still outperformed because they could see the bigger truth: the whole thing needed a risky rewrite, not patches.
The shape of work gets weirder: more pull requests, more review, new agent operators
As non-technical people gain the ability to ship technical changes, the volume of pull requests and half-finished output explodes. Shipper says that creates new roles for forward-deployed engineers — people who spend their time tuning, supervising, and systematizing agents so the rest of the company can operate at a higher level. He resists calling it babysitting; to him it's a fascinating systems job that turns AI mess into organizational leverage.
AI writing gets normalized faster than people expect
One of his sneakiest predictions is that we'll read a lot more AI-generated documents and emails — and often prefer them. He describes using a Notion agent for quarterly planning and letting Codex draft emails so well that one message sent to an investor without review was exactly what he would have written. His standard isn't 'no AI writing'; it's whether the human actually stands behind every line.
The big winners: PMs, full-stack designers, and anyone who 'rides the model'
Shipper is especially bullish on PMs, using Every's Marcus — an Axios PM turned AI-native builder — as proof that strong product taste plus light technical fluency can now ship absurdly fast. He's equally excited about designers who can finally turn taste into implementation instead of throwing specs over the wall. His closing advice is simple and practical: ride the models, keep turning over rocks, and find a use case that's genuinely fun, because the edge of AI isn't San Francisco — it's wherever a real human meets a new model and discovers what it's actually good for.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
