Why cultivating agency matters more than cultivating skills in the AI era | Max Schoening (Head of Product, Notion)
Lenny Rachitsky, Max Schoening · podcast · 2026-05-09 Reader 🖥️
AgencyMalleable SoftwareNon-Engineers BuildingRiding Model ImprovementsSmall Teams & CompressionTaste & Judgment
In the AI era, the differentiating factor for product builders is not skill acquisition but agency — the willingness to treat the world as malleable and to make things — and the products that endure are those built around a tiny, exceptionally good core rather than an accumulation of features.
practical · mixed — A long-form podcast conversation that is primarily practical — arguing for what product builders should do and how they should orient themselves in the AI era. It mixes expository segments (how Notion operates, tool choices) with argumentative segments (agency over skills, malleable software, the SaaS apocalypse being exaggerated) and narrative segments (anecdotes about GitHub, Heroku, Notion designers coding). The practical thrust: cultivate agency and taste through making, not through role definitions.
Structure
| Part | Role |
|---|---|
| Designers and PMs coding at Notion: the playground origin story | Concrete evidence that role boundaries are dissolving; establishes credibility for the thesis that skills matter less than willingness to engage with the material |
| Agency as the key differentiator | Core thesis segment — argues that with AI providing skills on demand, what separates people is agency: the belief that the world is malleable and can be changed |
| Malleable software as a design philosophy | Extends the agency thesis from individuals to products — software should work closer to the interest of users than the corporation, just as agency means individuals reshape their environment |
| The SaaS apocalypse is exaggerated | Counterargument to the claim that AI means everyone builds their own tools; defends 'as a service' as maintenance and specialist thinking people will always pay for |
| The first 10% of every project is now free | Describes how AI has changed the economics of exploration — cheaper to try many paths, demos not memos |
| Intelligence saturation and the retina display analogy | Argues that smarter models may not be the bottleneck; speed and modality may matter more past a threshold of 'good enough' |
| Taste as trained prediction: reps and feedback loops | Defines taste as the ability to predict how an in-group will react to an idea, built through iterative making — connects to agency through craft |
| The tiny core thesis: what makes great products | Central product philosophy — all great products have one tiny superpower; adding features never rescues a weak core |
| Quality, craft, and the missing engineering layer | Critique of current vibe-coding culture — quantity of software has increased but quality has not; the industry is neglecting the 'factory optimization' phase |
| UBI hot take and closing advice | Provocative framing that knowledge work is already UBI; advice to not let AI frenzy distract from what you care about; exercise agency by walking through the built world and realizing you could remake it |
Key terms
- agency Agency
The internalized belief that the world around you is malleable and that you can change it — not organizational maneuvering but the disposition to make things and reshape your environment. Schoening contrasts it sharply with skills. - malleable software Malleable Software
Software that works closer to the interest of the people who use it than the corporation that makes it — users can modify behavior, not just configure settings. Positioned between locked-down apps and run-your-own-Linux. - the tiny core Small Teams & Compression
The single, exceptionally good capability at the heart of a great product — multi-touch for iPhone, the pull request for GitHub, blocks for Notion, git-push for Heroku. The thing that makes everything else flow. - playground Non-Engineers Building
A separate, LLM-friendly codebase optimized for low-fear prototyping by non-engineers — designed to be 'oneshotable' so designers and PMs only have to overcome the fear of the terminal - obviously good Taste & Judgment
A quality bar defined by immediate recognition — 'you know it when you see it,' like the first iPhone or ChatGPT. Used as an internal Notion standard ('Let's just only make obviously good stuff'). - taste Taste & Judgment
The ability to run a virtual machine in your head where, given an idea, you can predict whether a specific in-group will like it. Built through iterative reps with feedback, not innate talent. Explicitly compared to training a model via backpropagation. - incremental correctness
Schoening's core value: iterate rapidly, converging on quality through many small steps rather than attempting to produce a perfect artifact in isolation - retina display analogy Riding Model Improvements
The idea that model intelligence has a saturation point — once you can't see the pixels, making them smaller doesn't help. Beyond 'good enough' intelligence, speed and modality matter more.
Propositions
- explicitcorroboratedAgency — the belief that the world is malleable and you can change it — is more important than skill acquisition in the AI era, because AI can supply skills on demand but cannot supply agency.conf: highsupports ←Building an Autonomous Enterprise for Real-World Services with Netic Founder Melisa Tokmak · p9 — Netic's hiring practice operationalizes exactly this claim, screening for demonstrated agency and follow-through over credentials or skills.
- explicitAgency is not evenly distributed in the world.conf: high
- explicitMaking things is the primary path to developing agency — tinkering awakens you to the realization that you can change the world around you.conf: high
- explicitThe value of designers and PMs coding is not shipping to production — it is forcing them to think in the medium and understand how agent loops work.conf: high
- explicitA designer or PM who understands how agent loops work is more valuable than one who can tweak UI styles in code.conf: high
- explicitThe first 10% of every project is now free — the cost of exploration has collapsed.conf: high
- explicitThe quality of software has not increased in the last 12 months despite a massive increase in the quantity of software produced.conf: high
- explicitAll great products have one tiny core that is a superpower; adding more features to a weak core never makes a product great.conf: high
- explicitcorroboratedModel intelligence will saturate for most knowledge-work tasks — once 'good enough,' speed, cost, and modality matter more than additional intelligence.conf: highsupports ←How the 4 New AI Models Change How You Work · p7 — Grok/SWE at a fraction of frontier cost plus Elon's 'most tasks don't require Fable-level capability' is a direct instance of Schoening's claim that once intelligence is 'good enough,' speed and cost matter more than additional intelligence.
- explicitSociety is largely not capped by intelligence.conf: medium
- explicitdevelopedThe SaaS apocalypse is greatly exaggerated — people don't actually want to maintain the full stack of software, and the 'as a service' part (maintenance and specialist thinking) retains its value.conf: highrefines ←The case for making your own apps · p5 — Where Schoening argues the SaaS apocalypse is broadly exaggerated because 'as a service' retains value, Mann's headless/records-holding-versus-weak-UI distinction refines that into a testable framework for which SaaS categories actually survive versus struggle.
- explicitSoftware tools will become more general (back toward word-processor/spreadsheet/FileMaker Pro generality) but will remain 'as a service.'conf: medium
- explicitcorroboratedTaste is the ability to predict how an in-group will react to an idea, and it is built through iterative reps with feedback — not innate talent.conf: highsupports ←Det skal være vanskelig · p11 — Schoening argues taste/judgment is built through iterative reps with feedback, not innate talent. This piece supports the mechanism empirically (those who best judged AI had done the tasks manually for years) and sharpens the bootstrapping tension Schoening leaves open: if AI does the reps, the judgment that would evaluate AI never gets built.extends ←The hidden pattern behind successful products | Mark Pincus (Founder of Zynga) · p5 — Shared claim: taste is judged by the target audience's reaction, not by peers'. Pincus's 'define your ambition in the eyes of your consumer, not your peers' specifies which in-group's reaction the taste-as-audience-prediction loop should optimize for.
- explicitcontestedTaste may not be an enduring human advantage over AI — the input-idea-to-reaction loop resembles backpropagation.conf: mediumcontradicts ←The best response to AI slop, infinite advice, and online noise is from Robin Williams · p13 — Schoening entertains that taste may not be an enduring human advantage over AI (the input-idea-to-reaction loop resembles backpropagation). Acunzo asserts the opposite conclusion — the human advantage is durable and located in non-reproducible embodied lived experience — without engaging the backprop mechanism; he relocates the advantage rather than defending its permanence.contradicts ←How Anthropic Builds And How Engineering Will Change Soon | Thariq Shihipar · p19 — Thariq treats taste as a durably valuable, increasingly important human differentiator; the other extraction argues taste may not be an enduring human advantage over AI since the idea-to-reaction feedback loop resembles a trainable process.contradicts ←Adam Mosseri: AI is a tailwind for authenticity · p2 — Mosseri claims taste/design is 'much more difficult to imagine being automated away' and is 'long on designers'; Schoening argues taste may NOT be an enduring human advantage because the input-idea-to-reaction loop resembles backpropagation. Direct disagreement on taste's durability.refines ←You can't unit test for taste · p4 — Schoening: taste may not be an enduring human advantage over AI. Tryggvason supplies evidence the LLM does carry latent 'taste' useful for ranking, but qualifies it — that taste is untrustworthy and must be human-verified (Reykjavik=10 doubt), so it augments rather than displaces human judgment.
- explicitAI's progress in non-coding domains (writing, etc.) has been far less impressive than in coding; what looks like domain progress is often just the application of coding principles to that domain.conf: high
- explicitEvery human intervention in code should feel like a bug in your development process — a signal that your verification loop has a gap.conf: high
- explicitYou don't have to be first to win — you have to be right. Durability matters more than virality.conf: high
- explicitKnowledge work is already a form of universal basic income.conf: medium
- explicitInclusivity is not always desirable — building for the top of the class in a domain sometimes requires being exclusive, and that's okay.conf: high
- explicitdevelopedThe merging of product roles risks losing specialists at the edges — the craftspeople who provide delight in design and reliability in engineering at scale.conf: highrefines ←Adam Mosseri: AI is a tailwind for authenticity · p7 — Schoening warns role-merging risks losing specialists at the edges; Mosseri supplies the mitigation — keep senior/creative specialists to pull in on demand and deliberately grow tomorrow's seniors.
- implicitProducts that succeed make their users feel like superheroes, rather than showcasing how good the product itself is.conf: high
- implicitNotion is better understood as an operating system than as a SaaS application — and the AI era makes this distinction strategically important because agents need a connected workspace to roam.conf: high
Arguments
Agency, not skill, is the differentiator in the AI era. strength: high
Explicit premises:
- AI models can now supply skills on demand — the 'insert skill issue' excuse no longer applies
- People who believe the world is malleable and act on that belief thrive; those who cling to role definitions struggle
- Agency is not evenly distributed
Implicit premises:
- Skills and agency are separable — you can have skills without agency (role-bound operators) or agency without skills (which AI now supplements)
- Agency is at least partly cultivable, not purely innate — otherwise the advice 'cultivate agency' would be vacuous
Inferences:
- Hiring should weight agency over demonstrated skills, especially for non-engineering roles
- If agency is unevenly distributed and increasingly the bottleneck, the inequality between high-agency and low-agency workers will widen in the AI era
Making things is the path to agency. strength: medium
Explicit premises:
- Tinkering and building reveal that the world is made by people no smarter than you (Steve Jobs quote)
- Making builds a treadmill of creating that leads to the realization you can change things
- Home-cooked meals, chairs, code — all forms of tinkering that develop agency
Implicit premises:
- Agency is fundamentally an epistemological shift — a change in what you believe is possible, triggered by direct experience of creation
- Passive consumption (of products, instructions, role definitions) reinforces learned helplessness
Inferences:
- Organizations that want more agency should create low-barrier making environments (like the Notion playground) rather than just exhorting people to 'be more agentic'
All great products are organized around a tiny, exceptionally good core. strength: high
Explicit premises:
- iPhone: multi-touch. GitHub: the pull request. Notion: blocks and slash commands. Heroku: git push. Dropbox: invisible sync.
- Adding features to a weak core never makes a product great — his own startup failed this way
Implicit premises:
- Product success is not primarily about breadth of functionality but about depth of one interaction pattern
- The tiny core creates an 'intoxicating' experience that generates enough goodwill to forgive everything else
Inferences:
- In the AI era, where the cost of adding features approaches zero, discipline about the core becomes even more critical — the temptation to add 'one more thing' will intensify
Model intelligence will saturate for most knowledge-work tasks, and after saturation, speed and modality matter more. strength: medium
Explicit premises:
- The retina display analogy: after you can't see the pixels, smaller pixels don't help
- Companies don't hire PhDs for every task — 'good enough' cognition is sufficient for most work
- Tyler Cowen's similar claim that society is not intelligence-capped
Implicit premises:
- The frontier labs are operating under the false assumption that users always want the smartest model
- The value function for intelligence in knowledge work has diminishing returns past a threshold
Inferences:
- Open-weight and locally-run models will become competitive for the majority of enterprise use cases, undermining lab pricing power
- The UX layer — how you interact with the model — will become the primary competitive surface, not model benchmarks
The SaaS apocalypse is exaggerated — 'as a service' retains its value even when creation costs collapse. strength: high
Explicit premises:
- People don't want to maintain the full stack of software, just as they don't want to hunt — they want the steak in styrofoam
- Software is a garden that needs tending (Brett Taylor)
- Slack notification delivery flowchart shows the depth of real-scale engineering
- Even Anthropic uses Slack rather than building their own
Implicit premises:
- The vibe-coding revolution lowers creation cost but not maintenance cost
- Maintenance and edge-case handling are the actual value of SaaS, not the initial feature set
Inferences:
- The SaaS companies most at risk are those whose value was purely 'a fancy form around a spreadsheet' — guiding users through a workflow that AI can now guide directly
Taste is trainable through reps, not innate, and resembles model training via backpropagation. strength: medium
Explicit premises:
- Taste means running a virtual machine predicting in-group reactions
- Japanese craftspeople build taste through decades of practice
- Designers with taste have side projects and are always tinkering with new apps
Implicit premises:
- If taste is trainable through the same loop as model training (input → prediction → feedback → update), then AI may eventually develop taste too
- The frequency of feedback loops matters more than their duration — increase reps, not years
Inferences:
- If taste is reducible to prediction-from-exposure, the claim that 'taste is the last human advantage' is weaker than commonly assumed
Narratives
- The Notion playground origin story: Schoening and two designers built a minimal, LLM-friendly codebase for prototyping chat interfaces because designing chat in Figma was 'drawing dead fish' (Brett Victor reference). This got designers on the treadmill of coding, and as models improved, they progressed to contributing to the production codebase.
→ Supports the proposition that the value of coding is mastering the medium, not shipping to production. Also demonstrates how to lower the agency barrier — make the first step non-scary. - Eric Lou's transformation: A PM who asked Schoening 'if you started a startup, would you hire me?' Schoening said 'not in the first 10.' Eric responded: 'I'm going to work on the skills so you'd hire me in the first five' — moving from writing strategy docs to Figma to building prototypes in code.
→ Exemplifies high agency — Eric redefined his role rather than waiting for permission. Supports the thesis that agency is the differentiator. - Brian Leven at Notion: blurs engineering and design, and is also Notion's top recruiter — not because it's his role, but because he identified what the org needed and went out to find people. 'Drive Notion like it's stolen.'
→ Exemplifies agency as going beyond role boundaries — not just cross-functional skills but taking ownership of organizational needs no one assigned you. - Schoening's failed Notion competitor (2014): Built a markdown editor with folding and all the features now in Obsidian. Spent too long polishing the editing experience. When Notion pivoted to document collaboration, True Ventures dropped his term sheet due to conflict. Meanwhile, Notion's first editor was terrible but the block model (tiny core) won.
→ Confirms the tiny-core thesis through a painful personal counter-example — working diligently on the wrong thing, adding features hoping to reach 'finally great,' while the competitor found the right core. - Dieter Rams walking through a museum, pointing at famous designer chairs with his walking stick saying 'this isn't good enough.' Schoening pins this to his Twitter profile because design should be first useful, then beautiful.
→ Supports the malleable software thesis and the jobs-to-be-done framing — products should be useful first, not museum pieces. Also connects to taste: Rams embodies accumulated taste from decades of reps. - Dropbox's menu bar icon was so good at syncing it was a better internet-connection indicator than macOS itself. Then Dropbox tried to expand surface area for years, and users kept wanting them to stay focused on the core job.
→ Supports the tiny-core thesis and the danger of feature accumulation — the best products do one thing so well that expanding away from it destroys value. - The Slack notification flowchart: the decision tree for delivering a single notification in Slack is enormously complex, representing decades of understanding real users at real scale.
→ Supports the argument that SaaS value lies in maintenance and deep customer understanding, not in the initial feature set that could be vibe-coded in a weekend. - Joanna Stern tweeted that thanks to Notion AI she finally understands and uses Notion. Schoening sees this as evidence that AI turns a hard-to-start general tool into one with a built-in tutor.
→ Supports the claim that software will trend toward general tools (word processor / spreadsheet generality) because AI removes the onboarding barrier that previously required SaaS to be narrow and guided.
Questions addressed
- ✅ What differentiates product builders who will thrive in the AI era from those who will fall behind?
Agency — the disposition to treat the world as malleable and to make things — is the differentiator. Skills can now be supplied by AI. - ✅ Should designers and PMs learn to code?
Yes, but not for the purpose of shipping to production — for the purpose of understanding the medium (especially agent loops). The prototyping use case matters more than the production use case. - ✅ Will SaaS survive the AI era?
Mostly yes — the 'as a service' part (maintenance, specialist thinking, scale engineering) retains value. What dies is SaaS that was merely a fancy form around a spreadsheet. - ✅ Will smarter models continue to be the primary driver of change?
Probably not — intelligence will saturate for most tasks. Speed, cost, local availability, and UX modality will become the competitive dimensions. - ✅ How do you build taste?
Through iterative reps with feedback — making things, getting reactions, updating your internal model. Frequency matters more than duration. - ✅ What makes a product great?
A tiny, exceptionally good core — not breadth of features. - ⭕ What do we lose as product roles merge?
Schoening identifies the risk (losing specialists, losing craft and delight) but does not offer a concrete mechanism for preserving specialization while encouraging role fluidity. - ⭕ Is taste the last durable human advantage over AI?
Schoening raises the question himself and admits uncertainty — if taste is just prediction-from-exposure (backpropagation), AI may develop it too.
Gaps
- incompleteAgency is presented as the key differentiator but its uneven distribution is noted without exploring why. Structural barriers to agency (economic precarity, organizational culture, psychological safety, access to tools) are absent from the analysis. The advice to 'just make things' assumes access and psychological conditions that are themselves unevenly distributed.
Read next: Amartya Sen's 'Development as Freedom' on capability theory; research on psychological safety (Amy Edmondson) as a precondition for agency in organizations - incompleteThe intelligence-saturation thesis (retina display analogy) is provocative but untested. Whether knowledge work has a clean saturation point or a long tail of edge cases requiring frontier intelligence is not examined. The analogy breaks if cognitive tasks are heterogeneous in their intelligence requirements.
Read next: Research on scaling laws and emergent capabilities in LLMs; Dario Amodei's 'Machines of Loving Grace' essay on where intelligence bottlenecks actually bind - incompleteThe 'tiny core' thesis for great products is well-illustrated but the episode does not address how you identify the right core before building. The examples are all retrospective — what made the iPhone's core obvious before launch? How do you distinguish 'tiny core not yet found' from 'no core exists'?
Read next: Marty Cagan on product discovery; the lean startup literature on validated learning; Clayton Christensen on identifying jobs to be done before buildingclosed byThe hidden pattern behind successful products | Mark Pincus (Founder of Zynga) — That gap asks how you identify the right 'tiny core' before building, given the examples are all retrospective. Pincus supplies a prospective heuristic method: isolate the gut instinct (the 'innovation zone') and hold it fixed; get a 'PhD in the proven' on the exact platform/audience; test many cheap ideas around the instinct, adding one 'better' that existing users unanimously endorse and one expendable 'new'; and start embarrassingly small so you can reach product-market fit before scaling. He also offers a heuristic for distinguishing 'core not yet found' from 'no core exists' — belief grounded in unmistakable signal ('lightning in a bottle') versus mere hope — while conceding (as this extraction's own gaps note) that the 'you'll know it's an A' test is itself fuzzy and not operationalized. - incompleteSchoening's claim that AI coding progress vastly outpaces other domains, and that what looks like domain progress is 'just applying coding principles,' is stated as an observation but not supported with evidence. It conflates the legibility of coding benchmarks with actual domain impact.
Read next: AlphaFold and AI in scientific discovery; AI in drug design (Recursion, Isomorphic Labs); progress in mathematical reasoning (AlphaProof)closed byLLMs Can't Jump — That gap asks for evidence on whether AI's coding-benchmark progress reflects real domain impact, naming 'progress in mathematical reasoning (AlphaProof)' as a specific reading direction. Zahavy engages AlphaProof directly and argues it demonstrates LLM mastery of Deduction only (formal derivation from given premises) -- not the Abduction/invention capability that would constitute genuine new-domain scientific progress. This complicates Schoening's 'domain progress is just applying coding principles' claim with a structural distinction rather than simply confirming or denying it; it does not address the gap's other named examples (AlphaFold, drug design), so treat as a partial close on the AlphaProof/math-reasoning strand specifically. - incompleteThe conversation repeatedly invokes 'quality' and 'craft' as things the industry is losing but offers no mechanism for measuring or incentivizing quality in an era where shipping speed is the dominant metric. The 'obviously good' standard is acknowledged to be subjective ('you know it when you see it').
Read next: John Ousterhout's 'A Philosophy of Software Design' on complexity as the enemy of quality; the Toyota Production System literature on building quality into the process rather than inspecting it afterward
Related documents
- Designing the AI-native engineering organization
- The founder of Postman says you have to kill your existing org chart
- Adam Mosseri: AI is a tailwind for authenticity
- AI Is Changing the Rules of Entrepreneurship
- The best response to AI slop, infinite advice, and online noise is from Robin Williams
- How Anthropic's product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)
- Addy Osmani’s Post
- The Founder's Playbook: Building An AI-Native Startup
- Bliki: Vibe Coding
- What is the harder human-capital problem beneath token capital?