Why cultivating agency matters more than cultivating skills in the AI era | Max Schoening (Notion)
In a Nutshell
In the AI era, agency—high initiative to experiment, prototype in code, and treat the world as malleable—outweighs specialized skills, enabling anyone to build the first 10% of projects cheaply and iterate via demos over memos. Successful products hinge on a tiny core superpower (e.g., Notion's blocks, iPhone's multi-touch), cultivated taste from reps, and malleable software like Notion that empowers users as creators rather than rigid SaaS consumers. Develop agency by making things, blurring roles (designers/PMs coding), and realizing "the world is made by people no smarter than you," while balancing rapid shipping with specialist engineering for scale and quality.
These notes were generated by AI and may contain inaccuracies.
Before, it was easy to say 'I will never be able to do this because [skill issue].' Now, even with skills at your fingertips, agency matters most. Agency is not evenly distributed. Advice for developing agency: 'Drive Notion like it's stolen.' One day you realize the world is made by people no smarter than you, awakening the idea that you can change things.
Most changed in jobs: The first 10% of every project is now free; it takes almost no effort to build the first version of a startup. Taste means running a virtual machine in your head to predict if an in-group will like an idea. Develop taste through reps, like training a model.
Great products have one tiny core superpower that's exceptionally good. Biggest pitfall: 'If I just add one more thing to the product, it will be great.' That never works.
Max Schoening: Product manager at Google, ran design team at Heroku, design leader and engineer at GitHub under Nat Friedman, two-time founder, now head of product at Notion. One of the most successful AI-forward product leaders and deep thinker on AI's impact on building and using software.
Host excited about Max's unique insight as roles merge (designers becoming PMs, engineers, etc.). Max embodies this: hard to describe what he is; he's done all the things.
Broad question: What changes for product teams and building as AI integrates into workflows? At Notion, Max pushes designers and PMs to ship code.
Max joined Notion building chat interfaces designed in Figma. Referenced Brett Victor's talk 'Stop Drawing Dead Fish'—static chat images are dead fish; you must feel the AI.
Two designers and Max built a worst-possible playground: small codebase LLM-friendly, using tools LLMs excel at. Moved chat interface prototyping there to get started easily.
Playground: Separate area with AI tools vs. full Notion codebase. First version aligned with model capabilities then. Main codebase not always agent-friendly due to iterations and decade of patterns.
Optimized for least scary, most one-shot: overcome fear of terminal, then it's chatting. Recreated patterns and UIs from playground into main codebase.
Playground gets people on the treadmill. As models improve, same designers/PMs contribute to production codebase (lesser degree). Trend: work increases exponentially with model capabilities.
Very useful for designers to move from Figma to code. Always in 'designers should code' camp. At GitHub, pre-LLMs, designers were top contributors (10% to GitHub itself).
Now processes break: Designers code/prototype in code, but marketing asks to reverse-engineer into Figma for assets—silly busywork.
Issue with vibe coding: Software quantity increased, but quality/reliability hasn't much in last 12 months.
Not about pushing to production, but thinking/designing in the real medium engineers will use.
Doesn't matter if designers/PMs ship production code. Thinking in code forces considering the medium.
Prefer PM/designer who understands agent loops over one who tweaks UI/styles. Understand agent loops by building them in code (their material). Coding harnesses are like '90s operating systems.
Code for prototyping/mastering material, not just shipping features. Once awakened to material, lines blur to production code—but goal is mastery, not being a delivery cog.
What separates thrivers: Agency (always separated founders). Before: 'Can't do this due to skill issue.' Now with AGI-adjacent models, agency matters.
High-agency people see world as malleable. Low-agency stick to rigid roles: 'What does it mean to be a PM/designer/engineer?'
Notion employees ('Ninos') have above-average agency, especially designers.
-
Brian Leven: Drives Notion like it's stolen. Blurs engineering/design. Top recruiter: identifies org needs, talks to people, hires. Affects change beyond day-to-day.
-
Eric Lou: From strategy docs to: 'Would you hire me in a startup's first 10? No? I'll make myself first 5.' Led to Figma, then prototyping in code to show thinking.
Rethinking engineer role without memes. Risk: Lose specialists.
Software metaphors: Prototypes like 3D-printed enclosures (layer lines obvious). Engineering: Optimize factory for yield/precision at 100M-1B scale.
Discourse misses this—focuses on tokens/features shipped. Need engineering for scale.
Design: Quick usable UIs from design systems, but where's delight/craft? Don't lose edge specialists in role merging.
One day you wake up and you realize the world is made up by people no smarter than you.
Throughline: Making things. Tinker/create: chair tweak, home-cooked meal. Innately human (tool-making, art).
Not circumventing bad bosses—start making. People notice; awakens change potential.
Malleable software: Works to users' interests, not corporation's. Reject ivory-tower design (even as Apple fan).
Apps glue UI/data ownership—can't tweak great app slightly. Opposite: Run own Linux, but high maintenance.
Key: Ownership over computing life (decreasing). Build on platforms/OS encouraging malleability (e.g., Notion, with real-time collab/security).
Works with Jeffrey Lit at Notion on this.
Pinned Twitter video: Dieter Rams criticizing museum chairs (Frank Gehry-like). German: Funny, disapprovingly points with stick.
Design first useful, then beautiful. Chairs for display, not sitting—nonsense (e.g., 'neither orderly nor properly chaotic').
Ties to 'Timeless Way of Building'/Stewart Brand's 'How Buildings Learn': Best homes adapt over time, not architect-built.
Malleability tests usefulness: Tweak/change it.
Malleable software vs. SaaS: Build own tools, no need for Salesforce/Notion?
SaaS from the 2010s was a fancy form around a spreadsheet, guiding people to fill out forms less malleable than spreadsheets. The 'as a service' part matters because people don't want to maintain full software stacks. Rebuilding software like Notion in a weekend is frustrating; most prefer ready solutions. Software is like a garden needing maintenance, which SaaS provides through specialists. Tools will become more general, like Notion, which is malleable but should be more so. Notion AI acts as a tutor, helping users understand and build, moving away from traditional SaaS. Software will revert to 90s general tools (word processors, spreadsheets, FileMaker Pro) but as a service, with specialized tools for security. SaaS apocalypse is exaggerated; things won't stay the same.
Building products requires ongoing maintenance, feature additions, and feedback. Examples: Anthropic uses Slack as their OS despite capability to build their own; Workday is another. Specialization allows spending on tools like Notion instead of building internally. Time is better spent on core work like AGI than rebuilding Slack. Slack's notification delivery involves complex decision flowcharts from real user scale and decades of experience.
First 10% of projects are now free; no need for PRDs—build janky demos instead. Easier to create version 0.8 of a startup. First 90% is cheap, last 10% is 90% of effort. Explore multiple paths by sending 10 agents. At GitHub, emphasized 'demos not memos'—react to prototypes, changelogs, or blog posts. Builds iteration earlier, replacing waterfall.
Conflicted on plain text/markdown/code durability vs. chatting. Figma usage splits: some designers drop it, others love AI tools. Questions if direct manipulation goes away with agents. Automation vs. augmentation: fast models like Spark, Haiku, Opus fast enable instant inference vs. queuing jobs. Instant inference might restore direct manipulation, molding code like clay. Depends on model capabilities and intelligence saturation—like retina display, beyond a point need different modalities and speed, not more smarts. Labs assume everyone wants frontier models, but for knowledge work, good enough suffices—optimize for local, cheap, fast. Society not capped by intelligence; prefer exoskeleton over god in a box.
One PM is highest token spender across Notion (excluding auto security scanning/bug triaging). Policy is unlimited now to encourage exploration; in 6-12 months, companies will scrutinize ROI. Individual spends thousands to tens of thousands. Milestone: when token spend exceeds salary. Don't boast about token spend like lines of code—largest projects have few million lines. Meta faced flak for token leaderboard, but it prods people to enlist agents and build software factories. Overindex to change defaults like writing PRDs or running meetings same way.
Non-engineering roles adopt easily due to huge capability gap—intoxicating superpowers, then explain why not merge PRs. For engineering, minimize manual code intervention (except reviewing—review more). Human intervention feels like a bug in agent-filled processes.
Figma has two camps; Jevons paradox—usage may rise alongside AI coding. Dislike rivalry discourse. Terminal surprising: PMs in Cloud Code or CodeX use it, encouraged over GUIs to understand computing substrate. Designers using terminal. Conductor: mostly developer tools.
Coding improves exponentially; less so in other domains like writing (still AI slop). Software eating the world accelerates as software cost nears zero, embedding business practices in code. Software engineering permeates domains: HR automates without bugging engineers. Model progress applies coding principles elsewhere. All agents are coding agents building needed skills, like OpenAI o1. All harnesses resemble coding agents.
In 6 months, companies scrutinize costs. If lab-open weight gap widens, labs decide world (hates centralization). If not, diffusion to self-hosted models like Cursor, Intercom, Notion dabbling. Smaller models good enough, cheaper without lab margins. Lucky timeline: US has three competent labs (maybe four with Meta, aim for six/dozen). Like cloud wars: layers commoditize, avoid lock-in. Heroku vs. Kubernetes—Kubernetes won by empowering ops teams with choice vs. replacing them. Products that win make users feel like superheroes (Kathy Sierra). Code review tools fail by roasting code publicly vs. Claude/Codex where you publish 'you + AI' work.
Publishing work created with Claude gives bragging rights as a developer. Notion's AI agent is praised for being genuinely useful. It took a long time to launch. Notion has been at the forefront of AI; their first assistant launched before ChatGPT. Founders Ivan and Simon anticipated AI's impact. Every company claims to be AI native, but true integration matters. Notion's connected workspace provides rich context for agents, making a connected workspace valuable. Agents need context without narrow interfaces to data. Notion resembles an operating system like Unix for coding agents. Enterprise search handles automatic permissions. > "It was made for such a time as this." Notion has everything about the company, perfect context for AI. Malleable software has been around but was too hard until now.
Notion ships consistently like Anthropic. At GitHub and Notion, both had incredible first acts followed by identity crises. Remind teams they can just do stuff without preciousness. Users upset by lack of innovation more than breaks. Increase shots on goal through experimentation. Notion shipped new functionality last 6 months, but feature count is silly like lines of code. Prefer fewer great features with combinatorics. Struggling with software quality; even Labs has regressions and poor TUI frame rates. Desire Apple-like machined quality. Balance shipping and quality with obviously good stuff—you know it when you see it, like first iPhone or ChatGPT. Avoid isolation; use incremental correctness—iterate rapidly. Notion has six automation primitives; consolidate to core idea, even if delaying ships. Industry needs reconciliation of evolutionary branches, e.g., Claude desktop app's multiple tabs.
Taste predicts in-group reaction to ideas via mental virtual machine. Build through reps, iterations with feedback—like training models via backpropagation. No speedrun; Japanese craftspeople gain taste over time. High-taste designers do side projects end-to-end and tinker with new apps. Surround with tasteful things; Notion names conference rooms after icons like first typewriter, Macintosh, Porsche 911 to inspire better work.
Team building matters most. No single through line like best design or engineering. Avoid adding one more thing; great products have tiny core superpower:
- iPhone: multi-touch
- GitHub: pull request
- Notion: blocks and slash commands
- Figma: real-time collaboration
- Heroku:
git push heroku master(simple deployment to URL) - Dropbox: menu bar icon for perfect syncing (better internet indicator than Mac)
- Snapchat: disappearing photos
Being first overrated; must be right. AirPods succeeded despite not first. Durability over virality; build generational like IKEA. Anthropic started after OpenAI, less funding, now dominates—Dario succeeded twice.
Use to think holistically: what user hires product for vs. what you want. Zoom out as user, not employee. Example: landing pages use marketing speak, not friend explanation. Pretend whiteboard pitch matches design.
We already have UBI: knowledge work. Needs for contentment are less; we invent hierarchies. Humans insert into agent loops inventively. Knowledge workers in air-conditioned rooms are lucky. With AGI, would do same: code for intellectual challenge like chess/Go, despite machines better (sad Lee Sedol quit Go). Tinker, make world malleable—e.g., consider robotics.
Inclusivity not always great; small group theory—world run by group chats of 8 or fewer. Exclusivity valuable. Notion shouldn't aim for 8 billion users; top 500 million want different things than everyone. Everybody top of class at something.
Being okay with being exclusive sometimes is fine, as long as it's not like McDonald's with exclusive hiring practices as the only job in a location. Working in comfy air-conditioned jobs with the top of the class can be winning: build a really good product for them, which by definition excludes others. The TBPN guys exemplify this—they had 8,000 listeners, got acquired for hundreds of millions. They pitched: if we have millions listening, we've done something wrong. This was designed specifically for people in power in tech to influence and teach them, and it worked great.
Max doesn't think in win versus fail terms; he fails a lot every day. Big cultural annoyances include hiring mistakes in running teams. Early on, hiring designers that can code was challenging, so loosening that requirement led to a slippery slope. Better to have fewer polymath designers.
On the product side, GitHub Actions lacked good package management—they thought it wasn't needed. The world would be better if they'd thought about it harder.
In 2014, Max started a Notion competitor focused on website building. The week they were to get a term sheet from True Ventures, Notion pivoted to document collaboration, creating a conflict. True Ventures passed. They spent time polishing the editing experience with markdown folding (like in Obsidian now). Notion's first editor was terrible—all blocks, couldn't select between them—but it didn't matter. Lesson: working diligently on the wrong thing for too long is a huge fail.
When a product works, there's a tiny core thing that makes it amazing, drawing people back no matter how bad everything else is. They kept adding features in a death spiral, but the core wasn't good.
You can tell quickly if something is taking off—it's obviously good. But user studies might fall flat if users don't know how to use it. Key: don't give up on the core idea (80%), then relentlessly iterate the 20% until it clicks with users.
Silicon Valley is uncharacteristically full of people who don't love computers—they're there to make money, fearing the 'last train' or permanent underclass. This distracts from spending heartbeats on what you care about and are passionate in. Advice: don't let the frenzy distract you; it'll find a way. Work incredibly hard from 18-25, then less later. Avoid the frenetic worry of missing the last train—it's a hollow way to live. Zoom out, read history, read computer science history.
On jobs in the future: tune down the worry amplitude. History repeats more than it's novel. Tie it to agency—not needing certainty means you'll be fine. Extreme mindset of total change: no move shields you fully, and it's insular.
It's great to follow your passion... if it pays.
Recommended Books
- Code by Charles Petzold: The secret language of hardware and software. Surprising how many programmers don't know how computers work. No line of code until chapter 27.
- Tools for Conviviality by Ivan Illich: Contrasts tools enabling human ingenuity/autonomy vs. industrial-scale tools destructive to autonomy.
- Seeing Like a State: For executives creating systems. Warns against fake legibility neglecting reality (e.g., status reports don't reflect team reality). Humans crave signal over noise, but it's often less signal than thought.
Favorite Recent Movie/TV
Purposefully watches terrible movies to be entertained by unreal experiences—don't ask for recommendations. Liked Project Hail Mary book and adaptation; excites about humanity's future (jokes Notion OS empowers 5-8 people to explore galaxy). The Handmaid's Tale (late to it): Replace God with AI, and it hits close to home—heavy but good.
Favorite Recent Products
Curated list of beautiful products people should buy (taste emulation).
- Ghosty terminal emulator: Most use terrible terminals—use Ghosty.
- Moshi (M-O-S-H-I) for phone: Not free, well done; exploring it (codes on phone now).
- Corne (C-O-R-N-E) split keyboard: Open source; claw back agency in compute life—download schematics, build from scratch.
- Cevivi pocket knife: High quality, more expensive, but a good tool to have.
Favorite Life Motto
The universe is change and life is what you make it.
(Marcus Aurelius quote.) We cling to certainty, but there's none—live in the moment, enjoy the ride, no spoilers on how it ends.
Favorite German Words
- Tuftla: Like 'tinker' but less derogatory.
- Faba: 'User' emphasizing using up—highlights impermanence/wastefulness of products.
Max on X/Twitter (@max.dev—wants to be less addicted). Listeners can help: Go for a walk in a man-made environment, observe everything made by people no smarter than you. In 6-9 months, you could figure out how to make most things from scratch—realizing this gives massive agency. Exercise it.
Keep Lenny's Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
