Stripe Sees AI Agents Driving Most Web Transactions Within Three Years

Claude

Rigid infrastructural guardrails allow developers to safely deploy AI-generated code at scale, setting up an economy where software agents dominate high-frequency digital commerce.

AI-assisted summary of Patrick Collison on Claude Code at Stripe — original source. Check the original for full context.

Source published: 2026-09-24

Key ideas

Stripe achieved five-and-a-half nines (99.9995%) API reliability alongside rapid AI adoption, with one engineer merging over 600 AI-authored pull requests in six months with only a single revert. (2:11)
A team of two to three engineers built Stripe Projects and integrated 50 idiosyncratic third-party partner APIs in just eight weeks using Claude Code, realizing an estimated 6x engineering speedup. (6:34)
Stripe longitudinal data reveals that global new business formation has doubled year-over-year—outpacing the March 2020 e-commerce surge—with startups reaching $100,000 and $1,000,000 revenue milestones at elevated rates. (13:06)
Stripe predicts that within roughly three years, the numerical majority of digital transactions will be high-frequency, smaller-value payments conducted autonomously between AI agents. (16:33)

Detailed notes

## Background Published on YouTube by the official Claude channel on September 24, 2026, this interview features Stripe CEO Patrick Collison in conversation with host Boris. According to the video description, the discussion explores how Stripe builds with Claude Code, examining how isolated devboxes and rigorous infrastructure sustain five and a half nines of reliability even as roughly 36% of pull requests originate from prompts.

## TL;DR Stripe CEO Patrick Collison discusses how Stripe integrates Claude Code and its proprietary 'Minions' orchestration layer into engineering workflows without compromising 99.9995% API reliability. Backed by hard infrastructural invariants established years prior, Stripe has seen code quality rise, illustrated by an engineer merging over 600 AI-written pull requests with just a single revert and a small team delivering Stripe Projects at an estimated 6x speedup. Beyond internal developer patterns—such as fanning out exhaustive plans across 10 parallel devbox environments—Collison highlights macro platform data revealing that global startup formation has doubled year-over-year. Looking forward, Stripe is actively retooling its financial infrastructure for an imminent agent-to-agent economy where autonomous software agents execute the majority of online transactions.

## Atmosphere & style The conversation carries the pragmatic, high-bandwidth energy of two technical peers comparing notes from the frontier of software engineering. Rather than leaning on vague AI hype, the discussion is crisp and empirically grounded, balancing infrastructural rigor—canary deployments, mathematical invariants, and telemetry—with genuine intellectual wonder. The dynamic is effortless and collegial, moving from granular systems architecture to macroeconomic data, ending on an energized, playful note as both speakers realize how rapidly technical reality is outpacing their original agenda.

## Bulletin - [02:11](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=131s) Stripe achieved five-and-a-half nines (99.9995%) API reliability alongside rapid AI adoption, with one engineer merging over 600 AI-authored pull requests in six months with only a single revert. - [06:34](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=394s) A team of two to three engineers built Stripe Projects and integrated 50 idiosyncratic third-party partner APIs in just eight weeks using Claude Code, realizing an estimated 6x engineering speedup. - [13:06](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=786s) Stripe longitudinal data reveals that global new business formation has doubled year-over-year—outpacing the March 2020 e-commerce surge—with startups reaching $100,000 and $1,000,000 revenue milestones at elevated rates. - [16:33](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=993s) Stripe predicts that within roughly three years, the numerical majority of digital transactions will be high-frequency, smaller-value payments conducted autonomously between AI agents.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=0s) Stripe pairs continuous, multi-deploy-per-day velocity with five-and-a-half nines (99.9995%) API uptime by using ephemeral, instrumented devboxes and gradual canary rollouts. - [00:00](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=0s) Traditional monthly or quarterly deployment cadences in finance accumulate massive failure surfaces and isolate engineering teams from rapid customer feedback loops. - [02:11](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=131s) Data contradicts fears that AI generation degrades software health: 600 AI-written PRs merged by a single engineer in H1 yielded only one revert, while overall PR quality trended upward over 18 months. - [02:11](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=131s) Stripe's internal orchestration layer, 'Minions', provisions fresh VMs from Slack or web prompts, executing code changes, packaging PRs, and running comprehensive automated test suites. - [04:23](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=263s) AI agent safety is enforced through mathematical invariants and hard infrastructure boundaries—specifically semantic data tagging and granular permission boundaries built in 2017—rather than probabilistic alignment. - [04:23](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=263s) The long-term equilibrium of software security heavily favors defensive engineering over offensive exploits as AI enhances automated inspection harnesses and observability. - [06:34](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=394s) Claude Code is integrated into default devbox images at Stripe, serving as the direct operational foundation for building integrations and scaffolding products. - [09:10](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=550s) Rather than producing low-quality code volume, Stripe engineers redirect surplus productivity toward deep architectural refactoring and resolving deferred technical debt. - [10:05](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=605s) Existing codebases increasingly serve as prompts for entirely new architectures, reinforcing expectations that Stripe's overall code quality will be higher in three years. - [10:44](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=644s) High-leverage engineering is shifting toward the 'plan modality', where developers craft comprehensive architectural plans with frontier models and fan execution out across 10 parallel devboxes for multi-day tasks. - [10:44](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=644s) The friction between ideation and functional implementation has collapsed, eliminating developer side-project backlogs and multiplying returns to intellectual curiosity. - [13:06](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=786s) Stripe's platform data indicates global business creation per unit time is up roughly 2x, driven by small teams operating with unprecedented leverage. - [13:06](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=786s) Rising startup volume reflects strong economic quality rather than empty vanity metrics, as average revenue per new firm and rates of reaching $100K and $1M thresholds continue to rise. - [13:06](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=786s) Real-world formation metrics challenge the assumption of AI monopolization, pointing toward a decentralized economy powered by agile, highly productive small enterprises. - [16:33](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=993s) Stripe is redesigning its rails, CLI tooling, and integration protocols (such as MCP) to support autonomous software agents opening accounts and settling transactions independently. - [16:33](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=993s) Stripe's official outlook projects a future internet animated by a whirling vortex of frequent agent-to-agent microtransactions constituting the bulk of digital commerce within three years.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=0s) Personal Experiments and the Dual Mandate of Reliability and Velocity

Patrick Collison opens with a personal anecdote illustrating the power of current models: after installing a weather station at his home, he used Claude to build a custom multimodal model completely from scratch. The resulting system now predicts local afternoon weather with greater accuracy than the National Weather Service. This rapid, end-to-end agency sets the backdrop for a deeper discussion with Boris about how Stripe approaches remote environments, devboxes, and AI-driven development at scale.

Boris asks what gave Stripe the conviction to push heavily into isolated, remote development environments while many peers in the tech industry hesitated. Patrick explains that Stripe operates under two non-negotiable, competing pressures: maintaining "five and a half nines of reliability" (99.9995%) across its core payment APIs while preserving high-velocity continuous deployment. In traditional financial services, software is deployed once a month, once a quarter, or even once every six months. While that cadence creates an illusion of local stability, it comes at a punishing double cost: teams are cut off from immediate customer feedback, and major deployment milestones become terrifying, high-risk events weighed down by "this accumulated detritus of months of progress."

Stripe rejected that model because engineering teams want customer feedback multiple times per day. To achieve both extreme velocity and five-and-a-half nines of uptime, Stripe invested heavily in its development pipeline and quality assurance infrastructure. The foundation begins with ephemeral devboxes equipped with deep instrumentation, observability, and operational tracking, followed by canary rollouts that advance progressively from a handful of machines to 1% of traffic before going company-wide. Extreme reliability and extreme velocity, Patrick argues, cannot be traded off against each other; building rigorous end-to-end tooling is the only way to achieve both simultaneously.

- Stripe's core infrastructure targets five and a half nines (99.9995%) API reliability alongside continuous, multi-deploy-per-day production updates. - Traditional legacy finance release cycles (monthly, quarterly, or bi-annually) accumulate massive risk surface areas and eliminate fast feedback loops. - Progressive deployment at Stripe isolates failure domains: changes move from isolated devboxes to canary machines, expand to 1% of production traffic, and gradually scale upward.

---

[02:11](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=131s) 600 AI PRs, One Revert, and the Minions Orchestration Layer

The industry remains split on whether agentic code generation will serve as a tailwind or a headwind for system stability, with many fearing that AI-written code authored at breakneck speed will receive less human scrutiny and degrade software health. Patrick counters this skepticism with concrete internal data. On the morning of the interview, he spoke with a Stripe engineer who merged more than 600 pull requests in the first half of the year (H1). Every single PR was written using AI—and across those 600 merges, exactly one pull request had to be reverted.

```mermaid flowchart TD accTitle: Stripe Minions Agentic Workflow accDescr: Diagram illustrating the end-to-end flow of Minions orchestrating AI-driven development from prompt to deployment.

subgraph Trigger["1. Developer Intent"] A["Engineer Prompt<br/>(Slack or Web Interface)"] end

subgraph Minions["2. Isolated Orchestration"] B["Minions Orchestrator"] C["Spin Up Clean Devbox VM"] D["Execute Coding Tasks via AI"] B --> C --> D end

subgraph QA["3. Automated Verification"] E["Package & Build PR"] F["Run Full Test Suite"] G["Automated Scrutiny & Scanners"] E --> F --> G end

subgraph Deploy["4. Progressive Rollout"] H["Canary Rollout<br/>(Handful of hosts)"] I["1% Production Traffic"] J["Full Fleet Deployment"] H --> I --> J end

A --> B D --> E G --> H ```

To power this workflow, Stripe built "Minions"—an internal orchestration system layered on top of their VM infrastructure. Using Minions, an engineer can prompt from Slack or a web interface to request a specific feature or task. Minions spins up a fresh, dedicated virtual machine, executes all code changes, packages the pull request, and passes it through Stripe's automated test suite.

Stripe's empirical metrics over the past 18 months show that overall quality per pull request has actually increased. While incident rates per unit time have ticked up slightly, Patrick clarifies that these are almost exclusively low-severity issues caught early by secondary safety nets, leaving total system reliability essentially unchanged. Looking ahead, Patrick expects AI to make Stripe "quite a bit more reliable" over the next one to two years because LLMs enable novel observability, runtime instrumentation, and automated inspection harnesses that were previously impossible to build.

- In H1, a single Stripe engineer merged over 600 AI-authored PRs with only a single revert. - Minions allows developers to trigger end-to-end development workflows directly from Slack or a browser, provisioning throwaway VMs on demand. - Code quality per pull request has trended upward over an 18-month window. - Overall platform reliability remains rock-solid because automated secondary defenses detect minor regressions before customer-facing impact occurs.

---

[04:23](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=263s) Hard Barriers, Data Guardrails, and Defense-Advantaged AI

Boris asks for the concrete, "nuts and bolts" guardrails Stripe implements to prevent regressions and protect against risks like prompt injection as PR volume multiplies. Patrick stresses that software integrity cannot depend on probabilistic hopes or subjective scrutiny:

> "It does mostly come back to this idea of relying on invariants and hard barriers rather than things that are somewhat more subjective or discretionary or probabilistic. And so obviously it's nice if the model is aligned and is more likely to write correct code or secure code or whatever... You really do want to rely on guarantees."

Boris notes that encoding constraints directly into infrastructure—such as rigid data segregation so unauthorized boundaries are physically impossible to breach—is now paying massive dividends in the age of autonomous agents. Patrick fully agrees, revealing that Stripe began heavily architecting semantic guardrails, dataset tagging, and granular permission boundaries back in 2017. While this multi-year effort was originally designed to provide airtight internal security guarantees, it accidentally positioned Stripe perfectly for agentic coding: an AI agent simply cannot access or compromise systems beyond its strictly enforced perimeter. When asked whether AI will favor attackers or defenders, Patrick predicts that in equilibrium over the next few years, AI systems will be "substantially defense advantaging."

- Stripe avoids probabilistic safety assumptions, enforcing strict mathematical and infrastructural invariants. - A semantic data-tagging and granular permissioning framework initiated in 2017 created structural guardrails that constrain autonomous agents today. - Collison believes the medium-to-long-term equilibrium of AI capabilities heavily favors defensive engineering over offensive exploits.

---

[06:34](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=394s) Building Stripe Projects with Claude Code at 6x Speed

Claude Code is pre-installed across every developer box at Stripe, serving as the standard entry point whenever an engineer begins a new technical task. Patrick highlights the creation of "Stripe Projects" as a prime case study of this tooling in action. The product concept was directly inspired by developer workflows in Claude Code: when an engineer uses Claude Code to scaffold a non-trivial application, they inevitably need supporting third-party services—such as PostHog for product analytics, Vercel for hosting, or an external cloud database. Because many of these developer-tool companies already use Stripe for billing, Stripe realized it could coordinate with them to allow AI agents to instantiate and configure partner accounts automatically.

Conceived at the beginning of the year, Stripe Projects went from initial brainstorm to public launch in approximately two months with a team of just two to three engineers. During those eight weeks, that small group built internal APIs, orchestration harnesses, and integrations with approximately 50 external services, each with its own idiosyncratic API designs, rate limits, and edge-case bugs.

When Patrick asked one of the engineers what that project would have required before agentic tooling, the engineer estimated it would have taken a larger team at least six months. Assuming a team twice the size working for three times as long, that represents a ~6x speedup. Even when discounted for natural developer optimism, Patrick notes that if Stripe's thousands of engineers experience even a conservative 2x blended productivity gain across all initiatives, it represents a "preposterously large deal" for the company's output.

- Claude Code is baked directly into the default image of every Stripe devbox environment. - Stripe Projects allows AI agents to provision and configure third-party accounts (e.g., PostHog, Vercel, databases) seamlessly via Stripe. - Two to three engineers built the internal service architecture and integrated 50 idiosyncratic partner APIs in eight weeks. - The project achieved an estimated 6x efficiency multiplier compared to pre-AI engineering baselines.

---

[09:10](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=550s) Sessions Output and Refactoring Against "AI Slop"

When engineers are granted large productivity boosts, Boris wonders where the surplus time goes: do teams ship more projects or run more speculative experiments? Patrick confirms that Stripe is actively doing both. At the company's recent annual developer conference, Stripe Sessions, the volume of newly launched products and features dramatically exceeded any previous year in the company's history—far outpacing the modest head-count growth of the engineering organization over the same period.

Critically, Patrick addresses the widespread industry concern that AI coding tools will merely generate mountains of low-quality "AI slop." Rather than just cranking out disposable customer-facing code, Stripe engineers are actively redirecting their newfound efficiency toward deep architectural refactoring and codebase cleanup. Because generating, testing, and modernizing structural patterns is dramatically faster with Claude Code, teams are finally tackling long-deferred technical debt and system overhauls that previously felt too labor-intensive to prioritize.

[10:05](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=605s) Code Quality, Planning Modality, and Multi-Devbox Orchestration

Patrick Collison notes that in modern software engineering, "every codebase is now the prompt for another codebase," pointing to former Stripe engineer Jarred Sumner's work rewriting Bun as a prime illustration. Addressing whether artificial intelligence will degrade or enhance software engineering standards, Patrick firmly predicts that the overall code quality at Stripe will be distinctly higher in three years rather than lower. Boris agrees, pointing out the powerful "fluidity between tokens and infrastructure": while observers fixate on generating product code directly, some of the most potent applications involve spending tokens on intermediary layers, infrastructure, and automated guardrails.

When Boris asks about the internal organizational barriers Stripe encountered while scaling AI adoption, Patrick explains that adoption did not face resistance—it was met with immediate, widespread enthusiasm. Instead, the real inflection points stemmed from external breakthroughs: the introduction of Claude Code as a dedicated modality and the substantial model capability jumps of late 2025. The primary learning curve was cognitive—engineers had to build an intuition for what is now possible and fundamentally rethink their workflows.

Patrick highlights an emerging work pattern practiced by one Stripe engineer: rather than prompting iteratively on small tasks, the developer spends extensive upfront effort collaborating with the most capable frontier model to draft an exhaustive, highly detailed execution plan. The engineer then fans that plan out across 10 independent devboxes in parallel, running autonomous agents on separate components of the architecture for an entire day, or even multiple days, relying on the model's capacity for self-verification. Patrick admits that he personally undervalued this "plan modality" early on, but has since adjusted his own workflow as models advanced. Boris observes that his personal "graveyard" of abandoned side projects has evaporated because AI allows developers to finish everything before deciding whether to keep it; Patrick agrees, noting that "shower thoughts now get translated to actual existence at a far higher rate" and the practical returns to intellectual curiosity have surged.

```mermaid flowchart TD accTitle: Orchestrated Multi-Devbox Planning Workflow accDescr: Diagram illustrating an engineer designing an extensive architectural plan with a frontier model and dispatching it across ten isolated parallel devboxes for multi-day execution. subgraph PlanPhase ["1. High-Leverage Planning"] A["Engineer + Frontier Model"] --> B["Comprehensive Architectural Plan"] end subgraph ExecPhase ["2. Parallel Execution (Multi-Day)"] B --> C["Devbox 1 Agent"] B --> D["Devbox 2 Agent"] B --> E["Devboxes 3-9 Agents"] B --> F["Devbox 10 Agent"] end subgraph VerifyPhase ["3. Convergence"] C & D & E & F --> G["Agent Self-Verification & Integration"] end ```

- **Bun Case Study**: Jarred Sumner demonstrated the paradigm of using existing codebases as prompts for full-system rewrites. - **The 10-Devbox Pattern**: Moving beyond single-prompt coding to treating autonomous agents as a multi-day quantum of labor executed across parallel cloud sandbox environments. - **Lower Cost of Exploration**: The friction between ideation and working implementation has collapsed, making exploratory prototyping virtually costless.

---

[13:06](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=786s) Macro Data: Business Formation Up 2x in the AI Era

Boris pivots the conversation toward macroeconomic shifts, recounting his quarterly talks at Y Combinator. Where he once asked who used Claude Code (eventually seeing every hand rise), he now asks who writes 100% of their code using AI—and roughly 70% of founders at early-stage startups raise their hands. Boris wonders whether Stripe's bird's-eye view of commerce reflects this explosion in small, highly leveraged teams moving faster and tackling larger problems.

Patrick confirms this trend emphatically, explaining that Stripe possesses a 15-year longitudinal time series that serves as a dependable proxy for global business formation. While the pandemic in March 2020 drove a well-known surge in e-commerce, the acceleration Stripe has recorded over the past year is, on a relative basis, far larger than any inflection point in company history: new business creation on Stripe is up by roughly a factor of two per unit time. Because roughly 25% of all Delaware corporations incorporate via Stripe—and an even larger share of actual tech startups do—Stripe's visibility captures the broader market reality.

This boom is not confined to Silicon Valley; Patrick emphasizes that it is broad-based across international borders, often outpacing laggy governmental reporting. For instance, while official UK statistics published by Companies House showed a year-over-year decline in new business formation, research published on Stripe's economics Substack demonstrated an underlying entrepreneurial surge once platform activity was accounted for. Crucially, the average revenue per new business is increasing rather than diluting, and the proportion of startups crossing meaningful revenue milestones—such as $100,000 or $1,000,000—has risen substantially. For Patrick, this micro-level data challenges the prevailing narrative that AI will inevitably concentrate wealth into a centralized economy, pointing instead to a distributed renaissance of successful, high-velocity small firms.

- **2x Launch Velocity**: Rate of new commercial enterprises launching on Stripe per unit time has doubled year-over-year. - **Delaware Market Share**: About 25% of all new Delaware corporate filings originate through Stripe Atlas/Stripe infrastructure. - **Revenue Quality**: The surge is not just vanity projects; average revenue per new business is trending upward, with elevated rates of firms surpassing $100K and $1M milestones. - **Decentralization Counterweight**: Real-world incorporation and revenue data show entrepreneurship decentralizing rather than being monopolized by legacy incumbents.

---

[16:33](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=993s) Preparing Stripe for the Autonomous Agent-to-Agent Economy

Looking ahead, Boris asks how these shifts are reshaping Stripe's internal product roadmap. Patrick reveals that Stripe's product teams are actively preparing for an ecosystem where AI agents and Claude Code instances are the direct users of financial rails, acting as counterparties on both sides of a transaction. Stripe is interrogating fundamental assumptions about financial access: determining how an autonomous software agent signs up for Stripe, whether protocols like MCP (Model Context Protocol) are sufficient, how to ensure the entire Stripe API can be seamlessly orchestrated from the command-line interface (CLI), and which native currencies agents will select to settle debts.

Boris remarks that this constitutes a true "agent to agent economy." Patrick agrees, articulating the official Stripe house view: within approximately three years, the outright majority of digital transactions will take place directly between software agents. While high-value aggregate dollar volume may still flow through conventional enterprise channels, Patrick expects the internet to be animated by a "whirling vortex of reasonably small agent to agent payments." The interview wraps up with lighthearted banter, with Patrick expressing how much fun the discussion was, and Boris laughing that they only managed to get through about half of the questions he had prepared.

- **Autonomous Counterparties**: Designing workflows for scenarios where AI agents independently authenticate, hold balances, and issue invoices without direct human intervention. - **Interface Evolution**: Investigating MCP support, deep Stripe CLI parity, and automated developer tooling optimized specifically for agent orchestration. - **Three-Year Horizon**: Stripe anticipates that numerical transaction volume will become predominantly agent-to-agent within roughly three years, driven by high-frequency, smaller-value microtransactions.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=0s) this accumulated detritus of months of progress — Patrick Collison

> [02:11](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=131s) quite a bit more reliable — Patrick Collison

> [04:23](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=263s) It does mostly come back to this idea of relying on invariants and hard barriers rather than things that are somewhat more subjective or discretionary or probabilistic. And so obviously it's nice if the model is aligned and is more likely to write correct code or secure code or whatever... You really do want to rely on guarantees. — Patrick Collison

> [04:23](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=263s) substantially defense advantaging — Patrick Collison

> [06:34](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=394s) preposterously large deal — Patrick Collison

> [10:05](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=605s) every codebase is now the prompt for another codebase — Patrick Collison

> [10:05](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=605s) fluidity between tokens and infrastructure — Boris

> [10:05](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=605s) shower thoughts now get translated to actual existence at a far higher rate — Patrick Collison

> [16:33](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=993s) agent to agent economy — Boris

> [16:33](https://www.youtube.com/watch?v=S_lzYIvtEaQ&t=993s) whirling vortex of reasonably small agent to agent payments — Patrick Collison

## Mentioned Patrick Collison, Boris, Claude Code, Stripe, National Weather Service, Minions, PostHog, Vercel, Stripe Projects, Stripe Sessions, Jarred Sumner, Bun, Y Combinator, Companies House, Stripe Atlas, Model Context Protocol (MCP)

## Video description Stripe CEO Patrick Collison joined Boris to talk about how Stripe builds with Claude Code.

About 36% of Stripe's pull requests now start as a prompt, with Claude Code doing the work in an isolated dev box. Patrick explains how Stripe keeps five and a half nines of reliability, how one engineer merged 600 AI-written pull requests in six months with a single revert, and why he expects AI to raise the quality of Stripe's code.

Have a question? Let us know in the comments.

Learn more about Claude Code: https://claude.com/claude-code

Chapters

0:00 Claude-powered weather model and throwaway devboxes 2:11 600 AI-written pull requests, Minions, one revert 4:23 AI guardrails as infrastructure 6:34 Stripe Projects with Claude Code 9:10 AI code quality: every codebase is a prompt 10:44 Claude Code plan mode and 10 parallel devboxes 13:06 Stripe's data: new startups up 2x with AI 16:33 How AI agents will pay each other on Stripe

[Source](https://www.youtube.com/watch?v=S_lzYIvtEaQ)

中文简报

Stripe预计AI智能体将在三年内主导大多数网络交易严格的基础设施防护机制让开发者能够安全地大规模部署AI生成的代码，为软件智能体主导高频数字商务的经济模式奠定基础。 Stripe首席执行官Patrick Collison探讨了Stripe如何在不影响99.9995% API可靠性的情况下，将Claude Code及其专有的“Minions”编排层整合到工程工作流中。得益于多年前建立的坚实基础设施不变量支持，Stripe的代码质量有所提升；例如，一名工程师合并了600多个由AI编写的Pull Request，仅出现一次代码回退，而一个小团队以预估6倍的速度交付了Stripe Projects。除内部开发者模式（如跨10个并行devbox环境分发详尽计划）外，Collison还强调了宏观平台数据，显示全球初创企业成立数量同比翻倍。展望未来，Stripe正在积极重构其金融基础设施，以迎接即将到来的智能体对智能体（agent-to-agent）经济，届时自主软件智能体将执行大多数在线交易。 Stripe在迅速采用AI的同时实现了“五个半九”（99.9995%）的API可靠性，其中一名工程师在六个月内合并了600多个由AI编写的Pull Request，仅发生一次代码回退。 (2:11) 一个由两至三名工程师组成的团队利用Claude Code，在短短八周内构建了Stripe Projects并集成了50个各具特色的第三方合作伙伴API，工程交付速度预估提升了6倍。 (6:34) Stripe的纵向数据表明，全球新企业成立数量同比翻倍，增速超越2020年3月的电商激增期，初创企业达到10万美元和100万美元收入里程碑的速度也进一步提升。 (13:06) Stripe预测，在约三年内，数字交易在数量上的绝大多数将是由AI智能体之间自主进行的高频、小额支付。 (16:33)

Reading summary

Stripe CEO Patrick Collison contends that integrating autonomous coding agents into software development enables rapid engineering velocity and business formation without compromising system reliability, provided companies rely on strict infrastructural invariants rather than probabilistic safeguards. Balancing Continuous Deployment with High Reliability Targets (0:00) Stripe enforces a dual mandate requiring five and a half nines of API reliability (99.9995%) alongside continuous, multi-deploy-per-day production releases. Collison contrasts this with traditional financial services that deploy monthly or biannually, arguing that infrequent releases produce fragile rollouts burdened by accumulated code changes and disconnect developers from customer feedback. To maintain stability at high speed, Stripe isolates failures through ephemeral devboxes, deep observability, and progressive rollouts that expand from canary hosts to 1% of production traffic before fleet-wide deployment. Addressing industry skepticism regarding AI-generated code quality, Collison cites internal Stripe data from the first half of the year, during which an engineer merged over 600 AI-written pull requests with only a single revert. Stripe coordinates these workflows through an orchestration layer called Minions, which accepts engineer prompts from Slack or a browser, launches clean virtual machines, writes code, and runs full test suites. According to Collison, overall pull request quality has improved over 18 months, with secondary defenses catching minor regressions before customer impact. Enforcing Hard Architectural Invariants over Probabilistic Safety (4:23) To prevent regressions and vulnerabilities such as prompt injection as automated code volume scales, Collison argues that engineering teams cannot rely on subjective model alignment or discretionary human scrutiny. Instead, systems must depend on strict invariants and infrastructural guarantees. He emphasizes that models will remain probabilistic, making hard boundaries essential to prevent unauthorized execution. Stripe relies on a semantic data-tagging and granular permission framework initiated in 2017. Although originally architected for internal access control, the system established rigid barriers that now prevent autonomous coding agents from crossing security boundaries. In the broader security landscape, Collison forecasts that over the next few years, AI capabilities will prove substantially defense-advantaging by automating runtime inspection, instrumentation, and formal verification. Workflow Shifts from Direct Coding to High-Leverage Planning (6:34) With Claude Code pre-installed across Stripe devboxes, teams are reporting notable productivity multiples. In one case study, a team of two to three engineers built Stripe Projects—a system orchestrating accounts across roughly 50 external partner APIs—in eight weeks, representing an estimated 6x speedup over historical baselines. Collison notes that time saved is not merely channelled into shipping more features, as demonstrated at the Stripe Sessions developer conference, but also into deep architectural refactoring and reducing legacy technical debt. Rather than engaging in small iterative prompts, advanced developers at Stripe increasingly use an upfront "plan modality." An engineer collaborates with a frontier model to design an exhaustive architectural plan, which is then dispatched across as many as 10 isolated devboxes executing in parallel for an entire day or longer. The workflow relies on the models' self-verification against test suites, dramatically lowering the execution cost of turning speculative ideas into working prototypes. Macro Startup Formation and the Agent Transaction Economy (13:06) Stripe's longitudinal transaction and incorporation data—which covers roughly 25% of all new Delaware corporate filings—reflects a doubling in the rate of new commercial business creation per unit time. Collison highlights that this surge outpaces governmental statistical reporting, extends internationally, and is accompanied by higher average revenue per business and an increased share of firms reaching $100,000 and $1,000,000 revenue milestones. He interprets this as evidence of economic decentralization rather than incumbent consolidation. Internally, Stripe is adapting its product roadmap for an environment where autonomous agents act as direct counterparties on financial rails. Stripe is evaluating command-line interface parity, Model Context Protocol (MCP) integrations, and native settlement methods for non-human entities. Collison states Stripe's institutional forecast is that within approximately three years, the numerical majority of digital transactions will occur directly between autonomous software agents via high-frequency microtransactions. Collison's projections assume that frontier models combined with rigorous internal platforms can enhance codebase health rather than accumulate technical debt. An unresolved question is whether organizations lacking Stripe's dedicated orchestration platforms, progressive canary harnesses, and mature permission tagging can replicate these velocity gains without compromising operational stability.

Source quotes

[0:00] Patrick Collison: this accumulated detritus of months of progress [2:11] Patrick Collison: quite a bit more reliable [4:23] Patrick Collison: It does mostly come back to this idea of relying on invariants and hard barriers rather than things that are somewhat more subjective or discretionary or probabilistic. And so obviously it's nice if the model is aligned and is more likely to write correct code or secure code or whatever... You really do want to rely on guarantees. [4:23] Patrick Collison: substantially defense advantaging [6:34] Patrick Collison: preposterously large deal [10:05] Patrick Collison: every codebase is now the prompt for another codebase [10:05] Boris: fluidity between tokens and infrastructure [10:05] Patrick Collison: shower thoughts now get translated to actual existence at a far higher rate [16:33] Boris: agent to agent economy [16:33] Patrick Collison: whirling vortex of reasonably small agent to agent payments

Original transcript

[0:00] - But our house, we installed a weather station. [0:02] With Claude, I could design from scratch a multimodal model, [0:07] and it predicts afternoon weather [0:10] better than the National Weather Service. [0:12] (bright pleasant music) [0:24] - I think a lot of people in the industry, [0:25] they tried to move to Claude VMs. [0:27] They tried to move to these remote environments. [0:29] What gave you the conviction to do this? [0:31] - Well, reliability and security [0:33] are really important for us. [0:35] Stripe operates with five and a half nines of reliability, [0:38] so extreme reliability for our core APIs. [0:42] But at the same time, we want to be developing [0:45] and launching new products [0:45] and adding new features extremely quickly. [0:47] We want to have continuous deployment of our API. [0:50] And across financial services, [0:52] the way things typically work is [0:54] maybe things get deployed once a month, once a quarter, [0:58] certain places once a half. [0:59] And I mean, that provides a kind of local stability, [1:03] but of course at two enormous costs [1:05] when you don't get quick feedback from customers, [1:07] you can't ship things on a regular basis. [1:09] But then also it does mean that these moments of migration [1:12] are incredibly fraught and scary [1:14] because you've this accumulated detritus [1:16] of months of progress. [1:20] And so we though that was totally untenable. [1:21] In fact, when we're developing things, [1:23] we want feedback from our customers multiple times a day. [1:25] And so we really weren't willing to compromise on this. [1:28] And so in order to achieve [1:30] five and a half nines of reliability, [1:32] but with this continuous process [1:33] of feature evolution development, [1:35] we though we really needed to invest [1:37] in the end-to-end development and quality assurance process. [1:42] And so the devboxes are just the first part of that, [1:44] and we can have some instrumentation there and observability [1:46] and operational characteristics and so forth. [1:49] And then a whole process of incremental deployment [1:52] where you first roll it out to, you know, [1:54] a handful of machines and then 1% of machines [1:57] and some progressive rollout strategy from there. [2:00] But I guess the general reason for our conviction, [2:04] we needed extreme reliability and security, [2:08] extreme velocity, [2:09] and I think it's the only way to achieve both. [2:10] - It's incredibly impressive, [2:12] the bar that you've been able to hit. [2:13] And I feel like everyone in the industry looks to Stripe [2:17] as the company that sets the bar for reliability. [2:19] - That's very nice to hear. [2:22] We obviously were working extremely hard to sustain that. [2:25] It's been interesting for us [2:26] in the era of agentic development [2:28] where obviously there's a question of, well, [2:30] will AI be a tailwind or a headwind here? [2:34] And there are obviously a lot of concerns of, [2:36] well, if people are writing code so much faster [2:38] and maybe they're not reviewing each line of the code [2:41] with quite as much scrutiny, lots of speculation [2:45] as to which way this will cut. [2:46] I was speaking with one engineer at Stripe this morning, [2:49] and over the course of H1, [2:51] he had more than 600 pull requests merged. [2:55] Every single one was written with AI, [2:57] and exactly one of those pull requests had to be reverted, [3:00] which at least suggests that it is possible [3:03] to have this enormously accelerated development rate [3:06] and with still empirically quite high reliability. [3:09] Minions of this thing on top of it, [3:11] which are a way to orchestrate VMs with prompts [3:15] from Slack or from a web interface [3:17] or, in principle, from any tool, [3:19] you can just ask for some feature to be implemented [3:22] or some task to be completed. [3:24] It will create a fully new VM for that [3:26] and then go and perform all the tasks along the way [3:28] and then it'll package it up and submit it and build it [3:31] and run it through our test suite [3:32] and the whole end-to-end process. [3:34] We have seen that quality per pull request [3:38] over the last 18 months has gone up. [3:42] Now, there's still some challenge [3:44] because the incidents per unit time [3:47] has gone up slightly. [3:49] And so now they're mostly very minor incidents [3:52] and all of our secondary mechanisms for catching things [3:55] before they become problematic have meant [3:57] that our total reliability is essentially unchanged, [4:00] which we take as quite heartening. [4:03] And then I think the cool thing is [4:04] obviously with LLMs and AI and everything else, [4:07] there's now of course the possibility to build [4:09] new observability and new instrumentation and new harnesses [4:13] and new kinds of automated scrutiny [4:15] that we couldn't build before. [4:16] And so I feel pretty confident [4:18] that over the next year or two that AI on net [4:21] will make Stripe quite a bit more reliable. [4:23] - So tell me a little bit more, like, nuts and bolts. [4:25] How do you ship faster using AI [4:29] without trading off against reliability and quality? [4:32] What are the specific guardrails that you have? [4:34] We wanna ship a model that's really aligned [4:36] and there's a bunch of protections, [4:37] there's a prompt injection protection. [4:39] And then also as the number of pull requests accumulates, [4:42] how do you make sure that everything remains reliable [4:45] and that code quality goes up? [4:46] - It does mostly come back to this idea of [4:51] relying on invariance and hard barriers rather than [4:58] things that are somewhat more subjective [5:00] or discretionary or probabilistic. [5:03] And so obviously it's nice if the model is aligned [5:07] and is more likely to write correct code [5:09] or secure code or whatever. [5:11] I mean, nothing's ever perfect. [5:12] You really do want to rely on guarantees. [5:15] - That's really interesting. [5:16] So what I'm hearing is, [5:17] before AI, there was a lot of benefits [5:20] to encoding guardrails and constraints as infrastructure. [5:22] - Yes. [5:23] - Because then engineers can make less mistakes. [5:25] And there's also benefits to data segregation [5:28] and making sure that people just literally [5:30] can't access data they shouldn't access. [5:32] And now this is kind of paying dividends. [5:34] - Yeah. I mean, this is full agreement with all of that. [5:36] And actually we started seriously investing a lot of this, [5:42] the apparatus and the tagging [5:45] and the sort of semantically aware guardrails [5:48] around our data back in 2017. [5:53] And it's been a multi-year journey [5:55] because there's a lot of annotating to do [5:57] and there's a lot of [5:58] granular permissioning to do and so forth. [6:00] And we did all of this [6:01] because we though the security guarantees were so important, [6:05] but you're right that it kind of accidentally [6:07] put us in a better position [6:08] when agentic development came along. [6:10] - Yeah, and then luckily Claude is great [6:12] at writing guardrails too. - Indeed, indeed. Yes. [6:14] Well, let's go back to the point where I really, [6:16] I mean, there's kind of a general question in the world [6:17] about whether AI and LLMs will be on net, [6:23] offense advantaging or defense advantaging. [6:26] But my guess is that in equilibrium in a couple of years, [6:30] this is all going to be substantially defense advantaging. [6:34] - I wanna come back a bit more [6:36] to how Stripe uses Claude Code. [6:38] Tell me about how do you use Claude? [6:40] What are some of the products that you've shipped? [6:42] - Every devbox has Claude Code pre-installed. [6:44] It's the first place people go when just, [6:47] the median person is going to complete some task. [6:49] It really has delivered meaningful acceleration. [6:52] One case that I though was quite informative here was, [6:55] we had the idea at the beginning of the year [6:56] of building a thing called Stripe Projects. [6:58] And this was kind of inspired by Claude Code [7:00] because when you're building some project with Claud Code, [7:05] for anything of any materiality or substance, [7:08] you invariably need to couple that [7:10] to some other set of services, right? [7:13] I want PostHog for logging or I wanna host it on Vercel [7:17] or I want a database, whatever. [7:19] Because almost all these companies [7:21] are Stripe customers already, [7:23] we thought we could work with them [7:25] to expose their capabilities in a new way [7:27] and make it incredibly easy for an agent [7:31] to instantiate an account with them. [7:33] We had this idea at the beginning of the year. [7:35] We decided to go and build it. [7:36] It was between two and three engineers [7:41] for about two months, [7:43] but from first idea to public launch. [7:47] And that's not only building the internal APIs and services [7:50] and harnesses and all the things, [7:53] but also integrating with now about 50 different services [7:58] and those services all have their own idiosyncrasies [8:01] and occasional bugs. [8:03] And I don't think you could have done that [8:05] with two to three people in eight weeks in the before times. [8:09] - What would that have taken before? [8:10] - So I asked one of the engineers involved [8:13] and his estimate was a bigger team, [8:16] didn't say exactly much bigger, [8:17] but a bigger team and six months. [8:19] Let's just say twice as big a team, I don't know. [8:21] And I guess that would've been 3x longer. [8:23] So that's a 6x relative change. [8:27] Again, maybe that's an underestimate, who knows, in that. [8:30] Software engineers are famously optimistic [8:33] when they estimate software projects. [8:35] So maybe we can say at least a 6x feed up, [8:38] which I mean Stripe has thousands of software engineers [8:40] and who knows if every project [8:42] is being accelerated by that magnitude. [8:44] I suspect there's a distribution. [8:47] Some things are massively faster, some things aren't. [8:49] But again, even we're really pessimistic if it's only, [8:52] let's just say it's only a 2x improvement [8:54] across the entirety of what we do, [8:56] I mean that's obviously still a preposterously large deal. [8:59] - Yeah, yeah, that's huge. [9:00] So it takes less engineers, they do it faster. [9:02] What do engineers do to- [9:03] - Again, with higher quality, at least per pull request. [9:06] - With higher quality, maintain the reliability bar. [9:09] So what do engineers do? [9:10] You just do more projects? You try more experiments? [9:14] - I think we definitely create more products [9:18] and try more experiments [9:20] and we can see this in the numbers. [9:23] We just had our annual conference sessions [9:26] and the number of new products [9:27] and the number of new features. [9:29] And again, the quantum of a feature is somewhat subjectively [9:34] and qualitatively defined, [9:35] but we try to be reasonably consistent year to year. [9:38] The number of new products and features [9:40] at sessions this year was wildly ahead of any prior year. [9:44] The engineering organization was somewhat larger, [9:46] but clearly there's some kind of productivity effect [9:50] happening there. [9:50] So we're certainly building more for customers. [9:52] But the other thing that I find kind of interesting [9:53] and that cuts against some of the slop concerns [9:57] is we're undertaking more projects now [9:59] to improve our architecture and improve our code base. [10:02] And the line from a friend that we often mention at Stripe [10:05] is that every code base is now the prompt [10:07] for another code base. [10:09] And of course, Jared Sumner, former Stripe, [10:12] has demonstrated this with the Bun rewrite. [10:16] My estimate of the... [10:18] There could be a question of, with AI and with LLMs, [10:21] should we expect the quality of code at Stripe [10:23] to be higher or lower in three years? [10:25] And I think the answer has to be higher. [10:26] - It's interesting. There's this sort of fluidity [10:28] between tokens and infrastructure. [10:30] You could use the tokens to write product code, [10:32] you can use the tokens to write infrastructure, [10:34] you can use it to write guardrails. [10:37] And I think people kind of over focus on using it [10:39] to directly build product, [10:40] but it's all these intermediary use cases [10:42] that are some of the most powerful. [10:44] What were some of the barriers that you hit getting AI [10:47] to this kind of scale internally? [10:50] - In general, it's been adopted very enthusiastically [10:55] and very completely and very quickly and so forth. [10:59] So in some sense, it feels funny [11:01] to ask out the barriers given the adoption fervor. [11:04] The phase changes were definitely [11:09] Claude Code itself as a modality. [11:11] So thank you for that. [11:13] And the model improvements of late last year, [11:18] another inflection point. [11:20] I think over the course of this year, I think the main thing [11:23] that people had to wrap their minds around is [11:26] getting a sense for what's possible [11:28] and how one can now work. [11:30] One person at Stripe, the way he likes to work is he will [11:35] work with the smartest model available [11:39] to construct a very complicated, extensive plan, [11:43] but really invest a lot in the planning process [11:46] and then maybe dispatch 10 different devboxes [11:50] with different agents executing different parts of the plan. [11:54] And they might work all day, [11:57] or in the extreme case, even for multiple days [12:01] on implementing this. [12:03] That's a very different way of working, right? [12:04] It's only very different to how things worked pre-AI, [12:07] but it's kind of non-obvious even given Claude Code [12:10] that, oh, I can, [12:12] the quantum of labor can actually be so extensive. [12:17] And again, if properly orchestrated [12:20] and guided that the model is capable [12:22] of such extensive self-verification. [12:25] I think that recently I myself undervalued the plan modality [12:30] and I guess goals have also [12:33] made a lot of progress over the course of this year [12:34] and I've changed how I use the model. [12:36] So anyway, I think there's just kind of getting intuition [12:38] for the ever changing and moving target of the modality. [12:41] - Yeah, I have this giant graveyard of these projects [12:44] that I started but never finished. [12:45] - Right. [12:46] - And that's just not really a thing anymore. [12:48] You kind of finish every project [12:49] and then you decide, do I want this or not? [12:51] - Yes, exactly. [12:52] And it's definitely the case that [12:55] shower thoughts now get translated to actual existence [13:00] at a far higher rate. [13:01] I think the returns to being curious [13:04] have gotten much higher. [13:05] - I wanna talk a little bit about what you're saying [13:08] on the Stripe side about how AI is changing business. [13:12] One of the things I've been thinking about is, [13:15] with AI, it is much easier for small teams [13:18] to compete against much larger teams [13:21] in a way that you just couldn't do before. [13:23] - Yep. - And I do these talks [13:24] at Y Combinator every three months or whatever, [13:27] and I used to ask, "Who uses Claude Code? Raise your hand." [13:30] And when we first released Claude Code, [13:32] a few hands went up, then at some point every hand goes up. [13:34] And now I ask a different question [13:35] 'cause that's not a useful question anymore. [13:37] I ask, who writes 100% of their code using AI? [13:41] And now it's roughly, I think like 70% of people write 100% [13:46] at these very small startups. [13:48] I think there's more startups than before. [13:51] They're moving more quickly than they were before. [13:54] They're tackling more ambitious problems. [13:55] - Yeah. [13:56] - I'm curious if you're seeing this in the data. [13:58] - Stripe launched 15 years ago, so we have [14:00] a decade and a half longitudinal [14:03] time series of firm creation, [14:05] but Stripe has been fairly popular among developers [14:09] for a reasonable number of years. [14:11] And so I think what we see in our time series [14:13] is some kind of proxy [14:14] for what's happening in the ecosystem overall. [14:16] And for example, during March of 2020, [14:19] we saw a significant acceleration [14:21] in new business creation on Stripe [14:23] for kind of obvious reasons. [14:25] The reason I bring this up is [14:26] because the acceleration that we've seen over the last year [14:29] is, on a relative basis, much larger [14:31] than any prior acceleration that we've ever seen. [14:35] The number of new businesses launching on Stripe [14:38] per unit time is up by roughly a factor of two. [14:41] And again, as kind of a one proxy metric for coverage, [14:45] about a quarter of all Delaware corporations [14:48] are now incorporated with Stripe. [14:50] And of the ones that aren't, [14:52] lots of them are, like, random subsidiaries for a, [14:54] you know, multinational or something. [14:55] So I think of the true startups, [14:58] it's an even larger fraction than 25%. [15:01] So I think we're seeing this pretty representative [15:03] and what we see is this huge acceleration. [15:06] And interestingly, the acceleration is, [15:08] it's quite broad based in the sense [15:10] that we see the same thing in most countries. [15:13] And in fact, in many cases, [15:15] government data hasn't even woken up to this yet. [15:18] And we recently published a piece in the Stripe economics, [15:20] Substack, about how the UK's official company statistics [15:24] show a year-over-year decline in new company creation. [15:27] Whereas in fact, and we kind of walk [15:29] through the methodology in the post, [15:30] we think the UK is seeing a huge surge [15:32] in new entrepreneurship. [15:33] And what we see is that the average revenue per new business [15:37] is going up, not down. [15:39] So it's way more companies [15:40] and the average outcome is improving. [15:43] And if you partition it [15:46] to look at businesses reaching 100K or a million [15:50] or again, any kind of arbitrary revenue threshold, [15:52] those are also all up very substantially. [15:56] And so I think it's just [16:00] a straightforwardly better time for entrepreneurship [16:04] than it was in the relatively recent past, [16:07] and that's pretty exciting. [16:08] And I think it's also one of the concerns with AI is that [16:14] it will bring about a more centralized economy. [16:17] But in the micro data that we observe, [16:20] we certainly see evidence for the opposite [16:23] where there is more new firm creation happening, [16:26] those firms are more likely to succeed [16:28] and growth rates overall are improving. [16:30] So we're pretty excited. [16:32] - Has this changed the kinds of products [16:34] that you're building? [16:35] Has it changed how you think about your own product? [16:37] - We're thinking about how are all the agents [16:40] and the Claude Code instances, [16:43] how are they gonna use Stripe directly themselves, right? [16:46] And both the Stripe product [16:48] and just how will they transact more broadly? [16:50] And so we're just thinking a lot about a world [16:52] where most transactions, [16:54] agents are the counterparties on both sides. [16:57] And so how does an agent sign up for Stripe? [16:59] How does an agent use Stripe? Is MCP enough? [17:03] What should the Stripe CLI do? [17:04] How to make sure that everything in Stripe can be [17:06] orchestrated and conducted from the CLI [17:10] or in some other way that is accessible to agents? [17:12] How will agents pay each other? What currency will they use? [17:14] - It's not just agent economy, [17:16] but it's like agent to agent economy. [17:17] - Yes. [17:18] Again, the Stripe house view is that [17:21] most transactions will be between agents within, [17:25] call it three years. [17:27] It may not be the case that [17:29] most of the dollar volume is directly between agents, [17:32] but I'll be sort of surprised [17:33] if there isn't just a whirling vortex [17:36] of reasonably small agent to agent payments. [17:39] - Patrick, thank you so much for coming by. [17:41] - Thank you for having me. That's super fun. [17:43] - Yeah, that was great. [17:43] I think I got to, like, half the questions. (chuckles) [17:45] (bright music)