Summary

Product development fundamentally changed in Dec ‘25–Jan ‘26 with long-horizon agents “resilient to failure.” Gokul’s tell: a video transcription tool he abandoned after debugging failures six months ago, he prompt-built in one hour while watching TV. Across portfolio CEOs, the big labs, and AI-native startups, the same shift: bottoms-up building, PMs committing code, prototyping interviews, and designer/PM-to-engineer ratios going from 1:3 to 1:20.
The one truly future-proof skill is judgment, because “you have the big challenge of AI slop” — “in an era when you can do everything, the question is which of these things matter.” The best product people are editors, not adders; Jack Dorsey called the PM “product editor.”
Thin AI applications face pressure. A Fortune 500 CIO: “I don’t know why I would use any of these startups” — he has Gemini’s agent builder, ChatGPT Enterprise, and 1,000 IT engineers wanting to be retrained. Meanwhile systems of record are cutting off APIs (Slack cut Glean), bundling free agents, or charging “$2 an API call” — so agent startups “have no option” but to build their own system of record plus multi-year migration tooling.
Software triage the public markets aren’t doing: seat/utility pricers (Zendesk) are most endangered — AI agents siphon 50 seats down to 20 as a two-way-door decision — and “many of them probably need to go private” to reprice on outcomes. Long-half-life data (NetSuite ERP, Salesforce records) is much more insulated: “it is career limiting to suddenly take NetSuite out.”
Ads: “Three and only three” ways to win — own coveted users on a first-party surface, deliver an outcome at a cost (AppLovin, “a 100-plus billion dollar company” on mobile installs alone), or be exclusive to big demand (Trade Desk/P&G). ChatGPT holds the dream hand: “their combination of intent and identity data is unparalleled.” Middlemen on the platforms and AEO shops “are not going to create durable enduring companies.”
What should scare incumbent ad networks: consumer behavior shifting to agentic interfaces they don’t own — the metric to watch is whether users who connect their Uber account to ChatGPT stop opening the app. For the new networks, being first doesn’t matter; Gemini could even position as “the zero ad platform.” Ads always tax engagement — run a no-ads holdout and set an explicit engagement budget, as Facebook did.
Patrick’s investor framing from 700 companies: explosive winners share high gross margins, low CAC, high retention, and a tight sales cycle — all downstream of true self-serve, which also makes products better (“the self-served customers were the most sophisticated users”).
Career call: become “a functional expert that knows how to build AI agents to do that function”; span of control under 10 “should not be allowed” — manage 50 humans or be an IC; and 12–18-month job hoppers are an “immediate red flag” because impact takes a minimum of 3–4 years.

Deep dive

1. Something fundamentally changed in December and January: the long-horizon agent

The anecdote Gokul builds everything on: six months ago he tried to build a video transcription tool with likely Claude Code — “it kept failing and then I had to go in and try to debug it. Ultimately, I gave up.” Two weeks ago, “while watching some episode of some TV show, in one hour I was able to basically prompt my way to a good video transcription tool because these agents now are resilient to failure.” He then surveyed portfolio CEOs, the large AI labs, and AI-native young startups — the same picture everywhere.
The org flipped bottoms-up: PMs now articulate customer needs at the highest level and act as “the guardian of the why,” while the product is built by engineers, researchers, PMs and designers working on the code together. PMs are checking code into production repositories — engineers still review, but soon likely Claude Code, Codex and other tools will review before commit — and interview loops have added an explicit “prototyping interview.” “If whatever you think of six months ago, if you continue thinking on that dimension, you have fallen behind.”
The headcount trade: teams offered an extra designer or an extra engineer are taking the engineer — design systems are already laid out and AI can design within them, leaving a small central team guarding the design language. Designer and PM roles are merging, and PM-to-engineer ratios are moving from 1:3 or 1:10 to 1:20.
Software went non-deterministic — do X, Y happens; do a slight variation of X, “completely something completely different happens” — so someone must own evals. “Who owns the evals? It’s the PMs” — and often “you got to write AI yourself to evaluate the results of AI because humans can’t.”

2. Judgment is the only future-proof skill — the best product people are editors

Patrick’s question — is anything truly future-proof at this pace? — gets a one-word answer: judgment. “Every product leader I’ve talked to is extremely worried that because you have these engines running rampant, they’re just going to produce lots of code” — AI slop. “In an era when you can do everything, the question is which of these things matter.” Critical code still needs human review: AI engineers can produce “beautiful code that could be wrong, that could have bugs in it, that could be vulnerable.”
The underlying philosophy: a product person balances customer needs and business needs, and no feature ships without a hypothesis “articulated in the form of a customer behavior change” — “we believe that by launching this thing the customers will go from doing X to doing Y.” Customer behaviors are the leading indicators of every business outcome.
On metrics: the north star “should not be revenue” — Square used GPV, Facebook moved from monthly to daily actives — and it must be paired with check metrics, because a team told to optimize one number will make it go up, full stop. DoorDash could grow GMV by setting delivery fees to zero and “the company’s revenue goes to zero” — so guardrails like gross margin or retention hold the line.
Jack Dorsey called the PM role the “product editor”: “any of us can look at a product and say here’s ten things you should build” — the job is cutting to the two that drive the customer outcome. Patrick’s echo: Rick Rubin said he wasn’t a producer but a “reducer.” Gokul: “Great example.”

3. Thin AI apps face pressure — the system-of-record wars force startups to build the whole platform

Where to build: agentic software “can do the job of people,” so target industries with highly paid, somewhat repetitive roles — “every three months the answer gets deeper.” Nine months ago nobody would have said a designer’s, architect’s, or lawyer’s job could be automated. But the bar is set by a Fortune 500 CIO he met: “I don’t know why I would use any of these startups. Gemini has an agent builder product… I also use likely ChatGPT Enterprise… and I have a thousand IT engineers” — all wanting to be retrained as AI engineers. If your customer’s CIO can build what you’re building, you lose.
The system-of-record counterattack: in 2025, some vertical incumbents began limiting access as agent companies (Salesforce; Epic in healthcare; Filevine and Clio in legal) stopped treating them like a “dumb database.” They’re blocking APIs — Slack, owned by Salesforce, cut off Glean — bundling their own agents free, or charging “$2 an API call or something”: “they’re trying to make the model of these agent companies unviable.”
So the system-of-action-on-top strategy may no longer be viable: “I don’t think that’s an option anymore.” Ambition must be to replace the entire platform — which means unglamorous migration tooling. One portfolio company “hired engineers in an Eastern European country for two years” just to build a Salesforce migration tool; at Square, small businesses wouldn’t switch POS even when cheaper because gift cards, loyalty and payments data were stuck. “Otherwise you’ve got to get Accenture.”
Durability checklist, since “the half-life of software today is so short”: a scarce asset (license, unique regulatory insight), a control point over money or data, hard-to-replace hardware, an essential workflow, or network effects — Harrison Helmer’s seven powers “embedded in the business model from day one.” Live examples: DoorDash’s restaurant–dasher–consumer network (Patrick: “you can’t vibe code your way to those two” — “Exactly”), money flowing through Toast or Mercury, Toast’s hardware, and Sierra’s scarce asset being Brett Taylor himself — “You can’t really outsell Brett.”

4. Software triage: seat pricing is endangered, data half-life is the moat

The most endangered software companies price on utility per seat. Zendesk: each seat is a customer-service agent handling tickets, so “I can have an AI agent sit right next to Zendesk” — pay for 20 seats instead of 50 with 30 AI agents alongside. The siphoning is gradual, “a two-way door decision,” not a rip-and-replace.
The fix is brutal: reprice from $20–30 per seat to “maybe a buck or 50 cents or 20 cents per ticket result” with no certainty how it turns out — “that’s why I think many of them probably need to go private” to make the business-model transformation off the public stage.
Insulation comes from data half-life. Slack is “precarious” because its data’s half-life is very short; likely NetSuite runs your whole business — “it is career limiting to suddenly take NetSuite out” — and the incumbent has time and data to train and bundle its own agents. The tradeable observation: “the software public markets are not distinguishing between these two types of companies.”

5. Superpowers shape companies in their image — and almost every founder or founding team needs an Eric

The pattern across Google, Facebook, Square and DoorDash: generational founders have “a superpower that is very aligned with what the company needs to succeed, and the company was really shaped in their image.” Larry and Sergey’s was technology and scale: the internal alpha of Caribou — Gmail, launched April 1, 2003 — offered 1 GB of storage when dominant Yahoo Mail gave 10 MB. Even with AdSense the fastest-growing product in Google history, Larry was disappointed it served under 1% of the internet’s ads — he wanted Google “involved in serving every single ad on the planet.” Same decade-long conviction behind Street View, TPUs, Waymo.
Zuck is “the greatest mind on building growth and engagement in consumer products” — and learns by shadowing. Within a year of sitting with the ads team he generated custom audiences, now the foundation of most ad systems, out of a likely Zynga complaint: likely Mark Pincus wanted whales, and Gokul said whales generated 80% of gaming-company revenue. Zuck asked “why can’t they just upload their whales into our system?… find them people similar to those whales.” It worked so well it generalized to all customer types. “He just has something about making connections between disparate domains.”
“Almost every great founder or founding team needs an Eric figure”: Zuckerberg had Cheryl Sandberg, Dorsey had likely Keith Rabois, Tony at DoorDash had Christopher Payne. Eric Schmidt’s 2007 strategy exercise: present every business unit with only images, no words — “People don’t remember words. They remember how things made them feel.” YouTube’s slide was just the hockey-stick of uploads per second since acquisition.
Once a company splits into “two rooms,” communicate deliberately: a weekly all-hands even at 15–20 people, plus a weekly CEO email with three sections — top of mind (spend 60–70% of the effort there), performance update, miscellaneous. “Don’t be afraid of repetition… that’s when it seeps into their bones.” And err candid: ask the team what they’d do — “people will rise up to the occasion.”

6. Lazy but brilliant: move risk from the gate to the transaction

“Square at its core is a risk company.” The founding story as told: Jack’s co-founder Jim, a successful St. Louis glassblower selling $2,000–3,000 sculptures, lost a sale when a woman calling from Panama couldn’t pay by card — banks had rejected him “many many times” for card acceptance. He and Dorsey realized the iPhone’s audio jack could take a reader. Where banks denied most small businesses, Square said “we are going to accept 95%” — and moved risk to the transaction level, scoring each payment with machine-learning models after onboarding.
Sergey did the same to AdSense in May 2003. The team had spent half its engineering on a publisher approval system; Sergey asked “why do you need to approve them?… What if they lie?” — anyone could apply claiming Nike.com. “We were just doing it to cover our asses, turns out.” His order: kill it, activate everyone instantly, and only start reviewing a URL after it hits 100 impressions. Click fraud, same posture: “How the hell do you solve that? The reality is you don’t. You just wait… and solve it when it needs to be solved.”
The principle — “lazy but brilliant onboarding”: “getting somebody to come to you and sign up is one of the rarest things in history,” so a pure self-serve product has zero upfront reviews and runs checks in real time on behavior. Jack’s design bar completes it: good design isn’t visually pleasing, it’s needing no manual — every point of sale requires days of barista training except Square, downloadable from the App Store.

7. Self-serve is both moat and classroom — and GTM now leads with outcomes

The self-serve religion also traces to Larry: shown Google’s internal customer systems built so sales and ops could manage big advertisers’ accounts, he ordered “end it right now… everything you’re building for large customers is also available to small customers.” The surprise: “the self-served customers were the most sophisticated users” — small agencies and hustlers “exploit the system in ways that you never even know,” and AdSense’s largest publishers signed up self-serve. He thinks Nike and Whole Foods signed up for Square devices.
The definition is strict — the customer onboards and uses the product “without ever talking to or engaging with a single member of the employee base” — which forces obsessive onboarding and a fast moment of delight, and opens the aperture: 100 salespeople reach maybe 10,000 customers; self-serve with word of mouth reaches millions. He bets Cursor is in every large company, yet “maybe 1% of companies is the top-down motion.”
His own Figma lesson in humility: after investing, he pushed Figma top-down into Square’s design team — designers refused, insisting Sketch was better, and he backed off. Two years later a mid-level design manager brought Figma in from a prior company and it kicked out Sketch. With self-serve “you can infiltrate and be an insurgent in a unique and powerful way which a direct sales motion could never have produced.” Patrick’s investor framing is that explosive winners share high gross margins, low CAC, high retention, tight sales cycles.
The GTM frontier: for consumer, scaling TikTok influencers — “somebody said TikTok is the best local search engine and I think that’s right”; his kids find restaurants Google Maps and Yelp won’t surface. For enterprise, outcome-based selling: likely Palantir’s pitch is “give us six months to solve it… If we can’t solve it, fire us. Don’t pay us anything.” “You cannot lead with what your product does anymore.” And go vertical for the lighthouse effect — land JP Morgan and every bank evaluates you; Proctor and Gamble means nothing to JP Morgan.

8. Ads: “three and only three” ways to win — and ChatGPT holds the dream hand

“As a company, you either die or you live long enough to become an ads company” — and it’s now happening to OpenAI. “There are three fundamental ways to succeed in the ads business. Three and only three”: own a coveted user group on your own first-party surface; deliver an outcome at a certain cost; or be the exclusive conduit for a large source of demand (Trade Desk taking Proctor and Gamble’s non-Google/Facebook display budget).
On the first path, “Google had intent data but not identity. Facebook identity but not intent.” likely ChatGPT has both: “their combination of intent and identity data is unparalleled… it’s the dream of any advertising person” — multi-phase natural-language sessions where Google typically loses the user after one click.
On the second, likely AppLovin — “a 100-plus billion dollar company” that drives one outcome, mobile app installs, and now “control[s] the buy side, the sell side, even the middleware… the auction for most mobile apps.” The doomed models: middlemen built on top of the platforms — “Google has the best engineers on the planet… they will take your capabilities, incorporate into their platform” — and the coming “cottage industry” of likely ChatGPT ad-optimizers and AEO shops, which “are not going to create durable enduring companies.”
For new entrants, “being first doesn’t matter” — first-party inventory isn’t going anywhere, and Gemini “doesn’t need to monetize anytime soon”; a strategic option is positioning as “the zero ad platform.” But holdout groups across many companies prove ads always cost engagement, so give the ads team an explicit engagement budget — Facebook set an annual cap on newsfeed engagement dip. And what should terrify Uber, Amazon, DoorDash: repeat transactions moving to agentic interfaces they don’t own — the signal is whether users who connect their Uber account to likely ChatGPT open the app less.

9. Hire doers, orchestrate agents, don’t job-hop

The number-one skill “probably even one year from now”: become “a functional expert that knows how to build AI agents to do that function and orchestrate an army of AI agents.” He cites a non-technical Meta PM who automated his own job so well “even his engineers [said] teach me how to use AI agents.” Management must be full-time: “span of control less than 10 should not be allowed at any company at this point” — manage 50 humans or be an IC. Company side: “don’t hire managers as long as possible. Hire doers.”
Assessment means work projects, because outside engineering “you can just BS your way without doing stuff.” Square’s corp-dev project: name one company Square should buy and why. “The best PM candidates rejected the premise completely” — one talked to ten Square merchants at Mint Plaza, found none wanted the proposed premium insights product, and pitched building something else instead. “You want agency.”
Tony’s DoorDash filter: hand candidates $10 or $20 and ask them to acquire 1,000 consumers. Nobody came close — one printed flyers at a gym — but it “was a brilliant way to just filter out people who didn’t want to do stuff.”
Career advice: “stay at every job long enough to have impact” — minimum three to four years. The 12–18-month “job optimizers” he’s seen over the last 18–24 months are “one of the biggest red flags as a hiring manager”; when he posted it on X, “tons of managers wrote to me saying it’s an immediate red flag” — and the rejected candidate “won’t even know what happened.”

10. Picking founders: authenticity, the idea maze, and boards as marriage

First question to every founder: “tell me your founding story.” He’s screening for authentic lived experience — Google, Facebook and DoorDash all started in school as near-toy problems the founders were curious about, and “starting a company because you want to start a company with your friend is the wrong reason.” likely Faire’s CEO Max Rhodes, who worked for him at Square, cycled through inauthentic ideas until landing on the one from his own life: as an undergrad running an umbrella company, he’d found getting local retail distribution brutally hard.
Second question: the idea maze — why this solution among the five or six others? He deliberately throws founders off-kilter with alternative approaches to test whether they’re “students of history in that industry.” The Collisons “bought a book on payments and studied exactly why all the payments companies did what they did and how they failed.”
Board hygiene: “a board role is like a marriage — once you get into it, it’s very hard to get out of.” Never seat anyone before spending a year with them on an advisory board. Pair every director with a management-team “board buddy” who meets or texts monthly between meetings — “it’s not the board meeting that truly matters. It’s all the things between the board meetings.”
The kindest thing: Bob McDonald hired him — “somewhat unqualified,” on a visa, never a PM — into one of the Valley’s hottest Sequoia-backed companies because he “saw a spark.” Gokul’s response has been to pay it forward with no expectation, grounded in gratitude: “we are sitting in literally the top 1% of the 1% of the 1% situations right now and breathing.”

Games Workshop: The World of Warhammer [Business Breakdowns Episode 239]AI & The Law: Changing Practice, Claude Constitution, & New Rights, w/ Kevin & Alan of Scaling Laws