Source: raw/Patrick_Collison_on_Claude_Code_at_Stripe.md — official Anthropic YouTube channel (@claude), interview with Patrick Collison (Stripe), youtube.com/watch?v=S_lzYIvtEaQ (fetched 2026-10-30). The Anthropic interviewer is not named in the transcript.
Collison explains how Stripe ships continuously with AI-written code while holding “five and a half nines of reliability” on its core APIs. Claude Code comes pre-installed on every devbox, and an internal orchestrator called Minions turns a prompt into a fresh VM, a built and tested change, and a submitted PR. The guardrails are hard infrastructure invariants rather than model alignment, and Stripe began building them in 2017 for security reasons. He also gives Stripe’s own data on new-business creation, which he says has accelerated more in the last year than at any point in Stripe’s 15-year history. All numbers are Collison’s, mostly from internal data or engineers’ estimates.
Key Takeaways
- Reliability and velocity together. Stripe runs its core APIs at “five and a half nines” and also wants continuous deployment, with customer feedback “multiple times a day.” Collison contrasts this with financial services, where deploys happen monthly or quarterly and each migration is “incredibly fraught and scary.” Stripe’s approach is to invest in the whole development and QA process: devboxes, instrumentation, then progressive rollout (a handful of machines, then 1%, then wider).
- Devboxes with Claude Code pre-installed. “Every devbox has Claude Code pre-installed. It’s the first place people go” when “the median person is going to complete some task.”
- Minions orchestrate whole tasks. Minions sit on top of the devboxes and “orchestrate VMs with prompts from Slack or from a web interface or, in principle, from any tool.” Each task gets “a fully new VM”, and the agent does the work, packages it, builds it, runs it through Stripe’s test suite and submits it.
- One engineer, 600+ AI-written PRs, one revert. An engineer told Collison he had more than 600 pull requests merged in H1, “every single one was written with AI, and exactly one of those pull requests had to be reverted.”
- Quality per PR is up; incidents per unit time are up slightly. Over 18 months, “quality per pull request… has gone up.” Incidents per unit time have risen slightly, but they are “mostly very minor”, and secondary catching mechanisms keep “total reliability… essentially unchanged.” He expects AI to make Stripe “quite a bit more reliable” within one to two years, because it also makes new observability, harnesses and automated scrutiny possible.
- Rely on invariants, not on the model behaving. “Relying on invariants and hard barriers rather than things that are somewhat more subjective or discretionary or probabilistic… You really do want to rely on guarantees.” Stripe started tagging data and building “semantically aware guardrails” in 2017, a multi-year job of annotation and granular permissioning done for security, which “accidentally put us in a better position when agentic development came along.”
- Stripe Projects: two to three engineers, eight weeks, about 50 integrations. Stripe Projects, inspired by Claude Code, lets an agent create accounts with other services (PostHog, Vercel, a database) through Stripe. It went from idea to public launch with “between two and three engineers for about two months”, including integrations with “about 50 different services.” An engineer estimated the old way would have taken “a bigger team and six months”; Collison’s rough sum is “at least a 6x speed up.”
- Even 2x would be huge. Stripe has “thousands of software engineers.” Speedups vary by project, but “even if… it’s only a 2x improvement across the entirety of what we do… that’s obviously still a preposterously large deal.”
- More products and more architecture work. The number of new products and features at Stripe Sessions this year “was wildly ahead of any prior year” (on a somewhat larger engineering org). Stripe is also taking on more projects to improve its architecture. A friend’s line Stripe repeats: “every code base is now the prompt for another code base”, with Jarred Sumner’s Bun rewrite as the example. Collison expects Stripe’s code quality to be higher in three years.
- A high-investment way of working. One Stripe engineer works with “the smartest model available to construct a very complicated, extensive plan,” then dispatches “10 different devboxes with different agents executing different parts of the plan,” running all day or for days. Collison says he had undervalued plan mode himself, and that “goals” have improved over the year.
- Firm creation on Stripe is about 2x. “The number of new businesses launching on Stripe per unit time is up by roughly a factor of two,” a larger relative acceleration than March 2020. About a quarter of all Delaware corporations are incorporated with Stripe. Average revenue per new business is going up, and so is the share passing revenue thresholds such as 100K or 1M. He says UK official statistics show a decline in company creation while Stripe’s data shows a surge (written up on Stripe’s economics Substack).
- Agents on both sides of a transaction. “The Stripe house view is that most transactions will be between agents within, call it, three years,” though not necessarily most dollar volume. Stripe is working on how an agent signs up for and uses Stripe, whether “MCP is enough”, and making everything doable from the Stripe CLI.
- Interviewer data point. The Anthropic interviewer says roughly 70% of founders at the YC talks he gives now write 100% of their code with AI.
Try It
- Give every agent task a fresh VM. Use a cloud session or a container per task, so each agent change starts clean and runs the full test suite before a human sees it (see Cloud Sessions Field Guide).
- Move one rule from the prompt into infrastructure. Pick a guardrail you currently enforce through instructions, such as “never read customer PII” or “never deploy without tests”, and enforce it with permissions, data tagging or CI so the model cannot get around it.
- Track the two numbers Stripe tracks. Revert rate per AI-written PR, and incidents per unit time. If quality per PR goes up while incidents per unit time creep up, add secondary catching layers, as Stripe did.
- Try the plan-then-fan-out pattern on one large task. Write an extensive plan with the strongest model in plan mode, split it into independent parts, and run each in its own session.
- Check how an agent signs up for your product. If your product requires a human sign-up, test whether an agent could create an account and use it through a CLI or MCP. Stripe expects most transactions to be agent-to-agent within about three years.
Related
- Stripe’s Kai — A Governance-First Company Agent — Stripe’s internal company agent, which lists Minions among its tools
- Agent Guardrails: Hooks, Permissions, and Sandboxing Patterns — the wiki map of hard-barrier guardrails
- Engineering Leverage in the Agent Era — Cherny’s case for encoding fixes as infrastructure
- How Anthropic Runs Large-Scale Code Migrations — includes the Bun rewrite Collison cites
- Claude Code in the Cloud: Field Guide to Cloud Sessions — the one-VM-per-task pattern
- goal Walkthrough — possibly the “goals” Collison mentions (unconfirmed)
- Factory — Building Autonomous Software With Droids — another case of planning with the strongest model
Open Questions
- The interviewer is not named. The YC-talks detail (“when we first released Claude Code”) suggests an Anthropic Claude Code lead, but the transcript does not say who.
- “Goals” in “goals have also made a lot of progress” is ambiguous: it could be Claude Code’s
/goalcommand or something internal to Stripe. - The cold open says Collison designed “from scratch a multimodal model” for a home weather station that “predicts afternoon weather better than the National Weather Service.” The interview does not explain how, or how it was evaluated.
- The “about a quarter of all Delaware corporations” figure is stated without a date or method; it presumably refers to Stripe Atlas, but the transcript does not name it.
- The 6x Stripe Projects speedup rests on one engineer’s estimate of the counterfactual (“a bigger team and six months”). Collison himself calls it rough.
- How Minions relate to Stripe’s Kai company agent is still not explained (see Kai).