Agency over skills: how AI is reshaping product teams

Executive overview

Skills are becoming table stakes. With AI handling execution, the scarce resource is agency — the belief that the world around you is malleable and that you can change it.

Max Schoening (Head of Product, Notion) argues that the designers and PMs who thrive are not the ones who can write the most code, but the ones who deeply understand the materials they work with and act without waiting for permission.

• Coding matters because it forces you to understand the medium, not because it ships features.
• The first 10% of any project is now free; the last 10% is still 90% of the effort.
• Software quality — not feature count — is the missing conversation in the AI era.

The world is made up of people no smarter than you. Start making things.

Agency as the new career differentiator

• Before AI, skill gaps were a credible excuse for inaction. That excuse is gone.
• Agency is unevenly distributed; people who treat the world as malleable will pull ahead.
• High-agency signals: changing your role to what it should be, recruiting outside your job description, building prototypes instead of writing PRDs.
• The treadmill metaphor: get people started on accessible tools; model capability growth does the rest.
• Coding is a forcing function — it makes you interrogate the material, not just describe it.

Malleable software and ownership

• Malleable software works closer to users' interests than the corporation's.
• Current apps are layered, glued systems — you can't tweak behavior the way you rearrange furniture.
• AI is awakening people to personal tool-making; that's malleable software in practice.
• The risk: individual tools in isolation don't benefit from collaboration or communal context.
• True malleability needs a platform that supports real-time collaboration and security — not just DIY scripts.

The SaaS apocalypse is exaggerated

• Much 2010s SaaS was a guided form around a spreadsheet — that layer is vulnerable.
• The as-a-service part is durable: maintenance, specialist expertise, and reliability don't go away.
• People don't want to hunt; they want the steak in styrofoam packaging.
• Software will trend back toward general tools (word processor, spreadsheet, database) but still delivered as a service.
• Specialized tools for security, compliance, and scale will persist — complexity doesn't vanish.

How AI has changed the work of building product

• The first 10% of every project is now free — no point writing a PRD when you can demo the janky version.
• Cheaper exploration: send 10 agents down 10 paths; review what comes back.
• "Demos, not memos" is now the default, not an aspiration.
• Waterfall is pointless — iteration is built in from day one.
• Quality hasn't improved proportionally to the volume of software being shipped; that's the gap to close.

Taste: what it is and how to build it

• Taste is the ability to run a virtual machine in your head — given an idea, predict how a specific in-group will react.
• You define the in-group; you don't need to build for 8 billion people.
• Build it through reps with feedback — identical to how models are trained.
• Designers with high taste maintain side projects where they own the full stack end to end.
• Surround yourself with excellent things so your own work always feels insufficient by comparison.

The tiny core: what makes great products

• Every great product has one tiny thing that is a superpower — multi-touch, the pull request, Git push Heroku master, Dropbox's menu bar icon.
• That core is found by stumbling and then the market confirming it.
• The death spiral: "if I just add one more feature, it'll finally be great." It never works.
• Being first is overrated; doing it right is what lasts.
• Incremental correctness — relentless iteration until the core clicks — is the method.

Roles, specialists, and what merging risks losing

• The danger in role-merging is losing specialists at the edges: engineers who optimize for a billion users, designers who bring delight and craft.
• The physical metaphor: a 3D-printed prototype is obviously not a manufactured product. Software has lost that distinction.
• Designers should think in code to understand the medium — whether that code ships to production is secondary.
• Understanding agent loops matters more than tweaking CSS; the former is the new core material.
• Every manual human intervention in a code-writing flow should feel slightly like a bug.

AI ROI and the coming reckoning

• Token spend is not a useful metric — same category error as bragging about lines of code written.
• In 6–12 months most companies will face uncomfortable ROI conversations about AI spend.
• If open-weight models stay competitive, fine-tuning on smaller, cheaper models becomes the rational choice.
• The cloud-wars parallel: businesses won't tolerate lock-in; commodity layers always emerge.
• The current moment — multiple frontier labs competing in the US — is unusually lucky. Don't assume it lasts.

Universal basic income and the knowledge work paradox

• Hot take: we already have UBI — it's called knowledge work.
• Humans will invent new reasons to insert themselves into agent loops; we always have.
• The real risk is letting the frenzy of "last train" thinking crowd out the work you actually care about.
• Agency and historical awareness are the antidote: history repeats more than it is novel.