Recorded 17 August 2026 · 54 min
Stripe's bet is that agentic coding should fund growth, not headcount cuts. Will Gaybrick says one engineer now does what two teams did two years ago. Stripe's response is, in his slightly cheeky phrase, to "build everything", because "the best way to optimize your cost structure is to grow more." The interviewer frames this as going long on future potential rather than going short. The evidence is Stripe's own network. The 2026 customer cohort is earning 50% more than the 2025 cohort, and new SaaS platforms are up 103%. Internally, minions now write about 30% of pull requests. The tension is that faster code creation moves the bottleneck to sellers, pricing pages and review. Gaybrick is also candid that agentic commerce is still at the missing-primitives stage. He calls the stablecoin case a possible "should work this way, may never work this way" problem. The conversation moves from product velocity to machine payments, stablecoins and tokens as a new kind of money.
Software now has costs — ▶ 2:28
Free-trial abuse was negligible when software margins were fat. Now compute is a real cost. Cursor was the first customer to feel it, with something like 1 in 6 free-trial users abusive: serial signups, even model distillation. Stripe stood up a pipeline in a weekend, combining its foundation model, network embeddings and a reasoning layer that says why an account looks like an abuser. ElevenLabs now blocks 2,000 free-trial abusers per day using Stripe signals. Identifying abusive actors across customers creates a network effect.
Win startups twice — ▶ 5:52
Strategy, as a Clerk founder summarized it: "win all the startups and then win them again." Startups are canaries for the next opportunity and grow into tomorrow's giants. They are also the toughest customers. Large enterprises call Stripe's reporting great because they are used to incumbents, while startups call it "garbage" and demand fixes. That pressure pulls Stripe up-market, and Fortune 500 companies, including Amazon and Microsoft, are now close to half of its customers. Enterprise differs in sales cycle and post-sale activation, but the principle of staying close to the user holds.
Each cohort beats the last — ▶ 7:58
First-half signups grew 50% year over year. The median 2026 cohort generates 50% more revenue than the comparable 2025 cohort, which itself beat the 2024 cohort by 70%. Two forces explain it. AI makes new products possible that could not exist four years ago, such as Suno and Higgsfield. And agentic coding means far fewer engineers are needed to build them. Billing usage grows faster than Stripe overall because it skews toward software creation. Software is expanding, not commoditizing.
Founder agency as the constraint — ▶ 11:13
With no one left to copy, the theory is that Stripe must become a platform for founders internally. Acquired founders already run Metronome and billing, crypto, open USD and Treasury. Peers cut opex after agentic gains. Stripe's view is to build everything against years of unmet user asks. The new constraints are back-office: code merging so fast it strains seller systems, pricing pages and seller training. The aim is to optimize the whole critical path from user ask to product in hand and strip out centralized process overhead.
Minions go one-shot — ▶ 14:22
Minions take a single prompt and go straight through CI/CD and testing to review, with no planning mode and no iteration. Stripe tracks the share of pull requests they write as a top internal metric. After the blog post in January or February they were doing 1,200 PRs per week. Last week it was about 7,000, roughly 30% of all PRs. Gaybrick says the one-shot, "glorified Ralph loop" approach is where the world is going, with agents given real power and the tooling made better and better.
Flatter, smaller teams — ▶ 15:55
Engineers have always been part engineer, part PM, part designer at Stripe, and that is intensifying. The shortest answer on organization is flatter, with fewer coordinating layers. Stripe Projects was built by a PM, a very senior engineer and a few others, with most PRs from one person orchestrating 16 agents. Gaybrick has no standard team size. Wider spans and smaller teams offset each other, so a line-managed team of eight does about 3x more with the same manager-to-IC ratio.
Going long, not short — ▶ 19:20
Kai, an internal knowledge tool built by two people, has 83% weekly and about 60% daily active use. Seller productivity is up 20%, and the conclusion is more sellers, not fewer, because payback improved. Gaybrick rejects "fewer engineers in the world" and invokes Jevons paradox. The new SaaS platform cohort is 103% larger in 2026 than 2025, despite a narrative that vertical SaaS would struggle. New grads become incubators of high-agency projects: spend management, adjacent to Treasury and roadmapped two years out, got picked up by one engineer. Near term, "resourcing back pressure" is healthy.
Injection molding for code — ▶ 24:00
Users wanted global tax filing after US filing took a long time. Stripe built global filing in about a third of the time, with much greater complexity. Gaybrick's analogy is injection molding: postwar consumer goods exploded once plastic could be melted and injected into molds. Stripe creates the molds. Agent files in every repo encode the pattern for integrating with financial institutions, then agents do the work. Remaining time goes to code review, where agents already review better than humans did. Timelines keep compressing, which makes growth a bigger lever than cost.
Missing primitives in agentic commerce — ▶ 28:02
Agentic commerce has had no Cambrian explosion and no repeated canonical use cases. Stripe's first diagnosis is missing primitives. With Tempo it built the machine payments protocol, where a service answers with a 402 saying how to buy it. Checkout for agents is still open: crawl forms or go native. Checkout pages should disappear even for humans, and Link's agent wallet and CLI let agents use credentials with humans in the loop. B2B looks strongest: Stripe Projects lets agents provision services like Vercel without a human visiting the site.
Micropayments finally make sense — ▶ 34:22
Microtransactions failed historically because content was squeezed between subscription and free. Agents change that. Gaybrick wants services to support ephemeral, one-time use so nobody opens a $9.99-a-month account for one song. Stablecoins make small payments feasible because agents handle the friction humans won't. Stripe has Stripe Treasury hold stablecoin balances, and covers about 60 countries in fiat but about 150 in stablecoins. Felix, built on stables, holds 5-10% of the US-Mexico remittance corridor. Stablecoins serve as the global Schelling point, though adoption may be a "should work this way" problem.
Tokens are blurring into dollars — ▶ 43:16
Gaybrick loves the word token because it sounds like an approximation of money. Attacks on general-purpose token consumers like Cursor and Replit resemble attempts to steal money from Stripe users, so the same mandate applies to tokens. As agents replace services, spend shifts from paying humans in dollars to paying fewer humans plus tokens. Stripe wants moving between tokens and dollars to be as safe as moving between dollars and euros, and to help customers manage tokens as both opex and product efficacy, such as choosing which model to use.
The cheapest way to read AI's effect on software is as a cost cut. Stripe is reading it as a growth lever. If one engineer can now do what two teams did two years ago, the company can shed people or it can clear a backlog. Will Gaybrick says Stripe is clearing the backlog. The interviewer's framing is long versus short. Cost optimization is finite and shorts your own future. Building more compounds. Gaybrick accepts that framing and puts it his own way: build everything.
The evidence he offers is Stripe's view of the whole market. The first-half signup count grew 50% year over year. The median 2026 cohort is earning 50% more than the 2025 cohort, which had beaten 2024's by 70%. Two things drive it. AI makes products possible that did not exist four years ago, and Gaybrick names Suno and Higgsfield. And far fewer engineers are needed to build them. Billing usage is growing faster than Stripe overall because it skews to software creators. A new SaaS platform cohort is 103% larger than last year's, in the face of a narrative that vertical SaaS would struggle. He does not buy the commoditization thesis either. Software is more than building. Abstractions must thread through the business, and expertise decays when the person who built a system leaves.
The same logic runs inside the company. Kai, an internal knowledge tool built by two people, has 83% weekly and about 60% daily use. Seller productivity is up 20%. Stripe's conclusion is not fewer sellers but more, because payback got better. Engineers face years of unmet user asks. Spend management, adjacent to Treasury and roadmapped two years out, was picked up by one engineer who just started building. Global tax filing took about a third of the time US filing did, with much greater complexity.
The mechanism is closer to manufacturing than to craft. Gaybrick compares agent engineering to injection molding. Stripe builds the molds: agent files in every repo that encode how it integrates with financial institutions. Then agents do the work. Minions are the purest version. A single prompt goes in, the code runs through CI/CD and testing, and a human reviews it. They went from 1,200 pull requests a week to about 7,000, roughly 30% of the total. One senior engineer orchestrating 16 agents produced most of the PRs for Stripe Projects. Teams get flatter. A line-managed team of eight does about 3x more with the same management ratio.
The obvious objection is that speed just moves the bottleneck. Gaybrick agrees. Code is merging so fast it strains seller systems, the pricing page and the ability to train sellers on new products. The binding constraints are back-office. Stripe's response is to optimize the whole critical path from user ask to product in hand, and to give everyone founder-like agency without the morass of centralized process. Quality is held the same way, from the top, repeated endlessly. Teams are also expected to use the product, including through simulated accounts that mimic a real user's disputes and refunds.
That is where the argument gets less certain. Gaybrick believes there will be more software engineers in the world, though fewer will be needed per task. He casts Stripe as an incubator for new grads. But he also says this is a good moment for "resourcing back pressure", because limits force efficient use of expensive new tools. He expects budgeting to look much as it did before. The evidence is Stripe's own cohorts, and Stripe sits on a long queue of unmet demand. Not every company has that queue. He offers no ideal team size and admits the forces of wider spans and smaller teams offset each other.
The same long-bet logic extends to what Stripe calls agentic commerce, but here the evidence is thinner. There has been no Cambrian moment and no canonical use case repeated over and over. Gaybrick says the first problem is missing primitives. With Tempo, Stripe built the machine payments protocol so a service can answer with a 402 stating how to buy it. What checkout should look like for agents is still open. Checkout pages should disappear, even for humans. B2B looks like the strongest case: agents can provision services such as Vercel without a person visiting a site. Consumers need something else, a way to use a service once without a monthly account. That requires microtransactions.
Micropayments failed before because content got squeezed between subscription and free. Agents change the buyer. They are hummingbirds, sipping a little data and a little compute from many places, and no human wants to open an account for each. Stablecoins make the payments cheap enough, and agents tolerate the friction humans will not. Stripe covers around 60 countries in fiat but about 150 in stablecoins. Felix, built on stables, already handles 5-10% of the US-Mexico remittance corridor. Gaybrick calls stablecoins the global shelling point, yet allows that the idea may be a "should work this way, may never work this way" problem.
Tokens extend the pattern. Gaybrick likes the word because it sounds like an approximation of money. Attacks on platforms like Cursor and Replit look like attacks on money. As agents replace services, spend shifts from dollars paid to humans toward tokens. Stripe sees a mandate to make the move between tokens and dollars as safe as moving between dollars and euros. What is unresolved is whether agentic commerce gets its canonical use cases soon enough to justify the primitives being built ahead of them.
The open question. The claim most worth testing is that falling engineering cost means more engineers, not fewer. Gaybrick says fewer will be needed per task but more will exist overall, and he backs it with Stripe's own customer cohorts and a backlog of unmet user asks. That is a company with a long queue of demand. He also concedes this is a moment for "resourcing back pressure" and expects budgeting to return to normal later. He gives no ideal team size and no sign of where the bottleneck lands once review and go-to-market are fixed. Watch whether the 30% minion share keeps climbing and whether hiring follows it.
Watch the full video at https://www.youtube.com/watch?v=P5iICDVn5gc.
Comments ()