The AI paradox: More automation, more humans, more work | Dan Shipper

Last time you dropped a spicy take that people were overlooking cloud code, especially for non‑engineering work, and you nailed it. Today we’re running through your next wave of predictions on how AI will reshape work, who thrives, and what to start doing now.

The job apocalypse story is overstated, and I’m very bullish on product managers and full‑stack designers. I’m deeply AI‑pilled and also very pro‑human, because every agent still needs a human steward and creativity is where we stand out.

You even doubled headcount this year at an AI‑forward company, which flips the script. Let’s set the stakes and revisit these predictions in a year to score them.

Our approach to forecasting is to live in the future together. Our team of early adopters gets early model access, we test everything, and we write about what we notice, which helps make it real.

Cloud code changed our workflow: we stopped staring at code and started talking to our computers in plain language, then used it beyond engineering for writing and operations. The simple test is whether you reach for it without thinking.

We’ll frame this in three buckets: how work changes, how the shape of work shifts, and who wins. Timeline‑wise, you expect the direction to be obvious within a year.

Work will bifurcate. You’ll delegate to a company super‑agent most likely in Slack, and you’ll do hands‑on work inside codex or cloud code on your own machine.

I flipped from personal agents to a single super‑agent per company because agents need an owner who cares. Without that gardener, they drift and break, so you start at the top and specialize over time.

So the near‑term reality is a shared agent in Slack that everyone can tap, with room for team agents later.

On the work surface, the big shift is an agent running on your computer with an in‑app browser, so it can see what you’re doing and act with you in one place. Codex’s desktop app nails this paradigm and has become my daily driver.

I keep a project thread, open a doc in our Proof editor, and let Codex watch and help. It also wrangles my inbox with our Cora agent, so I talk through emails and it researches, compiles documents, and sends the replies.

That implies SaaS runs inside codex or cloud code, not the other way around.

Exactly. Users bring their own tokens, so SaaS margins improve while you design products agents and humans can use together. You need clear HTML, a solid CLI, approval flows, logs, and fast rollbacks because agents move fast and in parallel.

Where does Cursor fit?

Cursor’s cloud setup is strong and they’re excellent for programmers, but they’ve chosen a narrower lane. Everyone now sees you need a harness above models, and the winning form is a surface that supports any kind of knowledge work.

For builders, the takeaway is to make products that agents can use directly and that humans can collaborate with in real time.

Right. Think joint human‑agent workflows where the agent drives the CLI and you see and steer in the UI. Simplicity matters because the agent can do the formatting and grunt work for you.

Many folks ran to the terminal. Do we snap back to GUIs with agents at our side?

We speed‑ran the terminal era. CLIs remain, but most work flows back to GUIs, especially for non‑programmers.

Big picture, we’ll chat with a shared Slack agent and do most computer work inside codex or cloud code, using apps within their internal browser.

And two agents often beat one. If you assume everyone works in codex or co‑work, onboarding flips: your agent shares rich context with the app’s agent, and setup and troubleshooting become almost automatic.

This also changes SaaS economics, since users pay for their model usage.

It saves margins and boosts demand. Agents grow the pool of SaaS users and volume, which is why I’m structurally bullish on SaaS even with new pricing and infra challenges.

You also argue automation adds work. What did you learn from your senior engineer benchmark?

Benchmarks can over‑signal autonomy. My test asked models to rebuild a vibe‑coded app from first principles; humans score in the high eighties to low nineties, and GPT 5.5 jumped to the low sixties because it finally shows the confidence to rewrite, not just patch.

But framing matters: models dutifully fix tickets, while a senior engineer says we should re‑architect and explains why. That judgment is hard to prompt and even harder to score, so humans stay in the loop.

And in the real world it’s humans using AI versus humans using AI, not humans versus AI.

Exactly. AI rarely uses itself in production; there’s always a person nearby.

Your quick playbook so far: work inside codex or cloud code, design your product for agent co‑use, and pilot a Slack agent for company knowledge.

And when you build for agents, sync the agent’s CLI actions with what people see in the web interface, because you’re collaborating on the same artifact.

How does the shape of work change as this takes hold?

Pull requests explode across the org as non‑technical folks ship code and content, which shifts the bottleneck to curation, coherence, and deletion. You need systems to decide what lands and how it fits together.

Roles blur as engineers design, PMs code, and marketers ship, which can feel destabilizing.

Generalists thrive, and a new role hardens: forward‑deployed engineers who manage agents day to day. They live in Slack, debug odd behavior, and keep agents useful, which is a durable job.

The result is faster shipping and a lot more reviewing; my data science friends now spend time vetting shaky analysis from others.

Solve that by building systems and bots so basic questions are answered reliably and experts focus on the deeper work. It’s less babysitting and more engineering a safe, scalable workflow for everyone else.

We’ll also read far more AI‑generated plans and emails and be fine with it when the sender owns the content. Our quarterly planning ran through Notion agents and produced strong drafts that leaders could refine fast.

Let’s talk who wins. You’re emphatic about PMs.

I am. A lightly technical PM with sharp product taste who rides cursor or cloud code can ship end‑to‑end and iterate with speed, which is devastatingly effective.

Music to a lot of ears.

Full‑stack designers are next. They can now encode the exact interaction they imagine, submit pull requests, and break out of the bland vibe‑coded look.

Meanwhile some roles, like sales, feel less changed, though top‑of‑funnel research and sourcing get a big lift.

The broader pattern is that models commoditize yesterday’s skill and humans push into what’s new. That is why I don’t expect mass unemployment, but I do expect people to adapt.

So how do we adapt right now?

Ride the models. Use codex or co‑work daily, try every new drop against your real work, and keep turning over rocks. The edge is wherever a real person applies AI to a real problem, and access is broad by design.

Stepping back, it’s wild: email, Slack, and SaaS remain, but every role morphs around them. Everything changes and nothing does.

That’s how it feels. Skip the doom and the utopia and focus on the next horizon; once you reach it, you find another one.

Lately it feels like big AI firms have dialed back the doom talk; scaring people about distant threats was starting to backfire.

That whole PR angle never made sense to me; even if it’s sincere, it’s ineffective and misguided.

Before we wrap, what should listeners actually do this year to keep up?

Run your daily workflows through a coding copilot and try an agent-style assistant, then keep what sticks; follow curiosity over fear and make it fun.

Find that first wow moment where AI saves you time, then build from there; ready for a quick lightning round?

We covered a lot, and I’m excited to see how my calls age in a year—please hold me to them.

Books you recommend the most?

Annie Dillard’s The Writing Life, especially the final chapter, is our house favorite. I also point people to Churchill’s history of World War Two for its builder-writer lens, and The Rigor of Angels for its Heisenberg, Borges, and Kant weave with striking AI echoes.

A recent movie or show you liked?

I’ve been into Knicks basketball and a mini-series called The Dark Wizard about Dean Potter, plus Hundred Foot Wave; I love intense pursuit stories that feel founder-adjacent.

A product you recently discovered and love?

Codex—it’s outstanding.

A favorite life motto?

Live stories worth telling, and write things people actually want to read; I also lean on a teaching from Robert Baya about meeting hard problems from spaciousness and strength.

An underrated AI tool right now?

Honestly, still Codex; paired with an in-app browser it handles email and analytics so well that it has reshaped how I work.

Can Anthropic catch up, and would you switch?

It’s a real horse race, and OpenAI feels back in favor after a rough stretch. I’ll always use the best tool, I’m not sponsored by anyone, and I still use Claude a lot.

Where can people find you and be helpful?

I’m on X at Dan Shipper; the most helpful thing is to have fun with AI, use it in your life, and share what works so we can learn together.

Dan, thanks for being here; if you enjoyed this, follow on Apple Podcasts, Spotify, or your favorite app, and a quick rating or review helps others find the show—past episodes are in the feed.