a16z Podcast

Will Gaybrick on why AI means build everything, not fewer people

Will Gaybrick· President of Product & Business at Stripe
·~54 min·English·a16z
AI CompanyAgentsBusiness Strategy
TL;DR

Stripe's Will Gaybrick argues AI is an expansion event, not an efficiency one: internal one-shot agents now write about 30% of Stripe's pull requests, teams are getting flatter, and the same abundance is reshaping agentic commerce, stablecoins, and the blurring line between tokens and dollars.

1Core Mental Model

Build Everything, Don't Cut

When one senior engineer can do what two teams did two years ago, most companies spent the gain on headcount cuts, but Stripe chose to spend it on building more, because cutting costs is finite and growing is not.

the best way to optimize your cost structure is to grow more.

— Will Gaybrick, a16z Podcast
Key Insight
This is Jevons paradox pointed at engineering labor: when each engineer gets cheaper to run, the rational move is not fewer engineers but more ambition, because Stripe still has years of unmet user asks to burn the surplus on.
2How They Build

Minions Ship 7,000 PRs a Week

Stripe's internal one-shot coding agents, called Minions, went from about 1,200 pull requests a week early in the year to roughly 7,000 a week, and now account for about 30% of all the code Stripe merges.

And last week, I think 7,000 PRs came from minions.

— Will Gaybrick, a16z Podcast
Key Insight
The tell is which metric Stripe watches: not lines of code but the share of PRs authored by agents. Managing to an agent-authorship ratio treats human review, not human typing, as the scarce resource, and implies that ratio only climbs from here.
3Org Design

A Platform for Founders, Turned Inward

Stripe has always been a platform for external founders, and Gaybrick now wants it to be one internally too: flatter, wider teams with fewer management layers, where a single senior engineer can orchestrate many agents at once.

if you want to ship more and build faster, you have to create founder-like agency inside your company.

— Will Gaybrick, a16z Podcast
Key Insight
Notice the org chart barely changes shape: the manager-to-IC ratio can stay the same while each team's scope triples. The gain shows up as throughput per person, not as a dramatic reorg, which is why Stripe resists a fixed ideal team size.
4Agent Engineering

Code by Injection Molding

Gaybrick compares building with agents to postwar injection molding: humans craft the reusable molds and templates once, agents mass-produce the code that fills them, and the scarce human time shifts almost entirely to code review.

that's sort of how we're doing code production at this point. you know, we're creating the molds and the templates

— Will Gaybrick, a16z Podcast
Key Insight
The valuable human skill migrates from writing code to designing the pattern the code must follow, which is why Gaybrick calls the markdown agents-files extremely important; a bad mold now mass-produces bad code at agent speed.
5Agentic Commerce

Commerce Is Missing Its Primitives

Agentic commerce hasn't had its breakout moment because the basic building blocks don't exist yet, so Stripe is trying to supply them, starting with a machine-payments protocol where a service answers a request with a 402 that says exactly how to pay.

you can request it and get a 402 response back. Um, and it just says like here's how you buy me.

— Will Gaybrick, a16z Podcast
Key Insight
The deeper claim is that checkout pages are skeuomorphic, a leftover human ritual; the native version lets an agent buy straight from a product page. That reframes investing too, since the test becomes whether an agent, not a human, would pick your service.
6The Agent Economy

Micropayments Finally Work

The old case against micropayments was that content gets squeezed between subscriptions and free ads, but when the buyer is an agent flitting across services, that logic breaks, and stablecoins give the agent a budget it can actually spend per use.

you want them to be these little sort of hummingbirds like going around the internet just slurping up a little data here

— Will Gaybrick, a16z Podcast
Key Insight
Micropayments failed for humans because the friction of a decision and an account outweighed a tiny purchase. Move the buying to an agent with a pre-set budget and the friction disappears, which is what finally makes per-use consumption economically real.
7Tokens & Dollars

Tokens Are Becoming Money

Platforms that sell tokens now get attacked the same way money accounts do, so Gaybrick sees the line between tokens and dollars blurring, and Stripe claims a mandate to protect and move tokens as safely as it does currency.

I've always loved the word token. Um, and the reason uh is that it just sounds like an approximation of money.

— Will Gaybrick, a16z Podcast
Key Insight
This is a quiet land grab. The fraud-prevention network Stripe built for dollars, spotting abuse across the whole payment graph, is exactly what token platforms like coding tools now need, so whoever secures token balances inherits that moat.
8Human Element

Taste Doesn't Vanish, You Simulate It

Asked how Stripe keeps its famous product taste when so much work is done by AI, Gaybrick's answer is a process, not a trait: make everyone use the product, and heavily simulate real usage with synthetic, PII-free accounts that feel alive.

we actually invest a lot in simulating the usage of products.

— Will Gaybrick, a16z Podcast
Key Insight
By making taste a manufacturable process, use it, simulate it, react to it, Gaybrick sidesteps the tired debate over whether models will ever have taste. If quality comes from disciplined exposure to real usage, it scales with tooling rather than depending on a few gifted individuals.