@TaylorPearsonMe @lennysan Great breakdown, glad you liked it!

The coding agent on your computer becomes a parallel work buddy. The surprise wasn't putting AI in a browser. It was putting a browser inside the agent that already lives on your computer and can see everything you do.

Dan has Codex watch him write, gather his email, run research, and execute tasks. He's been at inbox zero for 10 days straight (which he says never happens) because the agent collects every email and he just talks at each one.

This is what I'm calling freestyle work. You're not handing off a spec and waiting. You're riffing back and forth with something that can see your screen and that can act on your machine.

The one thing to do to keep your job: ride the models. The job apocalypse isn't really happening. A new model makes yesterday's human competence cheap.

Everyone can make a landing page now, so a default landing page is worth less and less. The people who stay ahead take the models and push into new, specific expertise the models haven't absorbed yet.

"Forward-deployed engineer" is becoming a real job: the human who runs the agents. Automation was supposed to delete jobs. Instead it created one. Every agent needs a human who cares about it. Dan's example: an engineer on his team spends most of his day in Slack talking to an internal agent that runs their consulting practice, mostly catching its dumb mistakes.

The inversion is the interesting part. The scarce, valuable role isn't being replaced by the agent. It's being the human in the loop the agent depends on. "Every agent needs a human" is going to be a whole category of work.

Company super-agents are beating personal agents, for now. Everyone assumed we'd each have our own agent. In practice they're brittle and need constant babysitting, so companies are converging on one super-agent for the whole company, maintained by a forward-deployed engineer. Shopify and Ramp already run this way.

The reason underneath it: an agent is only useful if a human cares about it and watches what it does. Sever that connection and it stops being useful. Which means the real product isn't the agent. It's the human who keeps it useful, and someone is going to build a service around exactly that.

Automation creates more human work, not less. Dan's prediction: as the models get better, you end up with more to do, not less. Automating a task spawns new coordination and management work above it.

Work splits in two: personal agents you delegate to, and a Codex/Cowork-style surface that becomes the operating system for all your knowledge work.

Automation doesn't subtract the work. It moves you up a layer, into direction and judgment.

The product managers who learn AI are going to win. Dan's example: Marcus, a PM who's "lightly technical" (knows what a database migration is, can read code if he has to), spent a year getting good at Cursor and Claude Code. Now he ships faster than almost anyone on the team, pairing modest technical chops with sharp product sense.

This is like the canonical freestyle example where a generalist with good process and AI tools is becoming as effective as a domain expert without them.

@TaylorPearsonMe I'm optimistic there will be more job creation than destruction. Niche's will go, but more adjacent ones created. The risk is ignoring it altogether and riding the model wave is a nice way of putting the solution.