TODAY’S POD SHOT

This is the third week running of the same underlying shape. Three weeks ago it was judgment. Last week, distribution. This week, agency. Different words, similar conversation..

❝

Hey there!

Remember, we've built an ever-growing library of our top podcast summaries. Whether you need a quick refresher, want to preview an episode, or need to get up to speed fast – we've got you covered. Check it out here

— Alastair

🎯 Agency Is The New Bottleneck

Skills aren't what's holding teams back anymore. That's the through-line from Schoening at Notion, Foroughi at AppLovin and Lütke at Shopify this week - capability is table stakes; what compounds is the willingness to act without waiting for the playbook.

Then Galloway shows up on Diary of a CEO and complicates it. He calls agency a privilege the financially cushioned already have, and argues AI is concentrating it further rather than democratising it. Hold both at once and "develop more agency" stops being a development plan. It becomes a question about what the company is actually funding.

Featured this week:

🕒 Estimated Reading Time: 11 mins. Time saved: 276+ mins! 🔥

Key insights from this week's brief:

  • 🏗️ Agency is the new bottleneck - Schoening: skills are table stakes; what separates builders is the disposition to act without permission.

  • ⚖️ Galloway's counterweight - Agency isn't a skill, it's a privilege. The rich already have it. AI concentrates it further.

  • 🧠 Taste is the remaining human moat - AI generates infinite options. Only humans with the reps know which one to ship.

  • 🚀 Foroughi runs £5B free cash flow with personal sign-off on every hire - One A-player with AI beats fifty average ones. He's betting the company on it.

  • 🔄 PMs are sending pull requests; nobody's built the guardrails - Role collapse is real. The wreckage is arriving on schedule.

  • 💡 The first 10% of any project is now free - When starting costs hit zero, the bottleneck moves to judgement about what's worth starting.

  • ⚡ The talent stratification trap is widening - A-players with AI compound at 100x. B-players get a 2x lift. The gap is opening, not closing.

⚡ Fault Lines

Three places the week's guests pull in different directions.

Skill or privilege? Schoening says agency is trainable. Galloway looks at the same landscape and sees class advantage - the people exhibiting agency in the AI era are largely the people who already had financial cushion, networks and runway. AI compounds that, it doesn't redistribute it. If both are partly right, "develop more agency" is asking employees to manufacture privately what was historically distributed by background.

Agency vs judgement. Singhal in Pod shots #132 said judgement was the moat AI couldn't take. Schoening this week says agency is. They overlap but aren't the same thing. Judgement is choosing what to do; agency is doing it without being asked. Product orgs that have one but not the other behave very differently in practice.

Liberation or chaos. Schoening reads role collapse as creative awakening. Mario and Armin on The Pragmatic Engineer read the same facts and see marketing pushing live changes, sales shipping un-reviewed features, PMs sending pull requests into code nobody owns. Same observation, opposite operating-model implications. Which one your engineering org is quietly running on matters more than most leadership teams have noticed.

Three weeks, three named answers, same underlying anxiety.

🚪 Not feeling this one? Try these from the vault

🏗️ The New Skill Is Agency

Max Schoening (VP Product, Notion) on Lenny Rachitsky's podcast defines agency as the disposition to act on the belief that the world is genuinely changeable by you, today. AI removed the technical excuse for not building. What's left is psychological, and Schoening's argument is that this is what teams should now hire and develop for - not portfolio depth.

Lütke on 20VC pushes the same point harder: whatever disposition someone walked in with, AI has amplified it rather than flattened it. The implication for product leaders is uncomfortable - the people who compound from here may not be the people with the strongest CVs. They may be the people who act before the playbook exists.

Key Takeaways:

  • The productivity unlock from AI is the willingness to act on new capability before the playbook exists, not technical fluency.

  • Hire and promote on observable agency, not portfolio depth. Portfolios are now a poor predictor of who will compound.

  • Agency inequality widens before it narrows. AI amplifies disposition; it doesn't distribute it.

  • The Steve Jobs realisation is now table-stakes context to articulate to your team, not a motivational poster.

⚖️ Agency As Privilege

Galloway pushes back on Schoening's framing on Diary of a CEO. AI wasn't built for the median worker - the people best placed to use it are the ones who already had capital, networks, and the financial runway to fail repeatedly until something hits. They had agency before AI. AI widens the gap rather than closing it. He's pointed about the tech leaders publicly catastrophising about AI safety while quietly buying back stock and shedding workers.

This doesn't make Schoening wrong, but it does make the prescription incomplete. If agency really is the new bottleneck and really is unevenly distributed, the organisations that benefit aren't the ones running agency workshops. They're the ones doing the structural work - psychological safety as a real budget line, genuine tolerance for failure, financial runway for ambitious bets that don't pay off immediately. The uncomfortable question for product leaders is whether the culture is building those conditions, or quietly selecting for people who arrived already insulated against the downside.

Key Takeaways:

  • Agency-as-virtue is convenient framing for orgs that don't want to fund the structural changes that make agency safe to exercise.

  • The AI productivity dividend flows disproportionately to people with the financial and network cushion to act audaciously. That's not a bug.

  • If your culture rewards agency for the already-confident and punishes it for everyone else, you're selecting for class background dressed up as merit.

  • Read Schoening and Galloway together. Either one alone is a dangerous half-truth.

🧠 Taste Is The Last Human Moat

If agency decides whether you act, taste decides whether you act on the right thing. Schoening describes it as a predictive model of human response - the ability to look at an idea and reliably guess how a target audience will react. It's built through years of reps that don't transfer easily and that AI can't replicate.

The real risk isn't PMs writing bad code with AI. It's the quieter erosion - product people outsourcing option-generation and option-selection to the model, and shipping work that's technically fine but never feels considered. AI is generous with breadth. The scarce thing is somebody who knows which tiny version of an idea is the one worth going deep on. Either the operating cadence creates the reps or it doesn't. Weekly product critique, public competitor reviews, "what we admired this week" - if nothing's scheduled, taste develops by accident, which usually means it doesn't develop at all.

Key Takeaways:

  • Taste is trainable but requires deliberate reps. AI doesn't shortcut the process and shouldn't be allowed to short-circuit it.

  • Great products are defined by depth on one tiny core, not breadth. AI generates options; taste selects the answer worth shipping.

  • Build a taste-development practice into the operating cadence: weekly product critique, "what we admired this week", public competitor reviews. If nothing's scheduled, you've outsourced taste to luck.

  • Protect specialist craft as a first-class budget line. Role collapse without specialists collapses into mediocrity fast.

🚀 The Lean-Plus-AI Bet

Adam Foroughi on the Founders Podcast describes AppLovin's 2022. The company IPO'd in April 2021 at around $28B, peaked near $40B, then watched its share price slide from $115 to $9 across the year - a 92% wipeout. Underneath, the business did over $1B of EBITDA in 2022, up from $700M the year before. The market called the company broken.

His response was a stock buyback rather than a hiring spree. Foroughi still personally approves every hire at AppLovin - the failure mode he's seen most often is companies that hire into opportunity, dilute their best people, and end up with a mess they can't unwind. The operating thesis is straightforward: in a world of B players, your 1x engineer with AI gets to 2x; your 10x engineer with AI gets to 100x. Lütke makes a similar point on 20VC about Shopify's best engineers writing far less code this year, not because they got worse but because they've moved a layer up. The implication for product leaders is the bit nobody wants to say out loud: "more people = more output" stops holding the moment AI becomes a real multiplier on the best people. Headcount stops being the lever. Calibre is.

Key Takeaways:

  • Hire for AI leverage, not headcount. One exceptional person with AI tools compounds faster than five average hires using the same stack.

  • Institute meaningful friction in hiring. Foroughi's personal sign-off model is a working pattern for preventing IQ dilution at scale.

  • The 100x multiplier requires AI access plus capability plus agency. All three. Two of the three doesn't compound.

  • Foroughi held the line on lean during 2022 even as EBITDA passed $1B and shares cratered to $9. His framing: "hire into opportunity, dilute their IQ, and you're stuck with a mess."

🔄 The Guardrail Crisis

Mario and Armin on The Pragmatic Engineer name something that's been building for a while - non-engineers participating directly in engineering processes. PMs sending pull requests. Marketing pushing live changes. Sales reps shipping small features no engineer has formally reviewed. The problem they flag isn't that any of this is happening. It's what isn't happening alongside it.

Schoening reads this as creative awakening; Mario and Armin worry about it. They're describing the same set of facts. The quality gates specialists used to own don't reconstitute themselves when roles flatten - they have to be rebuilt deliberately, and most organisations are deferring that work because the velocity uplift in the meantime feels worth it. The cost surfaces six to twelve weeks later, in production, at which point somebody starts asking who owns the bit nobody reviewed. "Everyone can ship" needs to come with "and here's who owns the standard," or the velocity gain pays itself back with interest later.

Key Takeaways:

  • Establish ownership of the quality gate before encouraging cross-functional code contribution. "Everyone can ship" without "someone owns reliability" is a recipe for production incidents.

  • The PM-as-builder model works when design and engineering are partners, not reviewers brought in after the fact.

  • AI-written specs from a prototype are first drafts, not final briefs. Engineering judgement still hardens them for scale.

  • Role collapse accelerates velocity but compounds quality regression. Budget process investment proportional to the velocity uplift.

💡 The First 10% Is Now Free

The first 10% of any project is essentially free now. Prototyping a feature is a morning's work rather than a sprint's. What changes is which question is actually hard. The interesting one is no longer "can we build it?" - that bar dropped. It's "should we even be starting this?", which is the question most product orgs have less practice with.

Chatbase shows what that looks like in practice. Yasser Elsaid took the company from first tweet to $1M ARR in 117 days with no sales team, and now serves customers from UN war-zone responders to theme parks. His next product is a "chief customer officer" - a single agent running support, sales and onboarding with full relationship context. AppLovin's Axon 2 runs the same closed-loop pattern on ad creative. Humans set the objective and check the output. The agent runs the loop in between. The teams getting genuine compounding from AI aren't treating it as a co-pilot, and they're rebuilding their workflows around the new shape. The question worth holding is whether there's a workflow on your team still running in co-pilot mode that should probably be running closed-loop instead.

Key Takeaways:

  • When starting costs collapse, the quality of your starting criteria becomes the scarce input. Invest there.

  • Self-serve plus closed-loop AI can outpace hiring sales, especially in markets where the product is the channel.

  • The first-10%-free principle applies to research and discovery as much as software. Watch for AI-native discovery engines in regulated industries this year.

  • The human role in closed-loop AI is objective-setting and quality gating. Teams still operating each step manually are competing on the wrong axis.

📋 Three Questions

  1. Agency as virtue vs. agency as privilege: Schoening says agency is the new bottleneck and is unevenly distributed. Galloway says it's not a skill, it's a class privilege AI is concentrating further. If both are right, your "develop more agency" mandate is asking employees to manufacture privately what was historically class-distributed. Is your hiring rubric measuring agency, or is it measuring confidence backed by financial cushion?

  2. Liberation vs. guardrails inside role collapse: PMs sending pull requests is either creative breakthrough or quiet quality crisis depending on which guardrails you've built. Which framing is your engineering org actually operating under, and does the rest of the company know?

  3. Stratification vs. Jevons paradox: Foroughi says A-players with AI compound 100x and B-players get 2x. The Jevons reading says total demand grows faster than productivity, so absolute headcount stays flat or rises. Both can be true at different scales. Does your three-year talent plan acknowledge that the macro and individual stories are different, or is it implicitly betting on only one?

🔗 If you liked this, here's where to go next

Five from the archive that extend the same thread:

That’s a wrap.

As always, the journey doesn't end here!

Please share and let us know what you liked or want changing! 🚀👋

Alastair 🍽️.

Reply

Avatar

or to participate