How Stripe's AI 'Minions' Ship 1,300 PRs Weekly from a Slack Emoji

When I first saw the Stripe "Minions" on my timeline, I knew I had to learn more. The idea that a team has built not just one agent, but a whole fleet of them to assist with daily development work is exactly the kind of practical, high-impact AI application we love to explore.

For over six years, Steve Kaliski, a software engineer from Stripe, has been at the heart of building developer tools and payment infrastructure at Stripe. He’s part of the team that brought these AI coding agents to life, and the results are staggering: they're currently landing about 1,300 pull requests per week with no human coding involved, just human review. Steve mentioned that he personally can't remember the last time he started his work in a text editor. Instead, his work begins where ideas happen—in a Slack thread, a Google Doc, or a ticket.

In our conversation, Steve walks us through two incredible workflows. First, we get a detailed, behind-the-scenes look at how any Stripe employee can kick off complex development work with a simple emoji reaction in Slack. This workflow dramatically lowers what Steve calls the "activation energy" required to turn an idea into code. Second, he demos a mind-bending future where AI agents act as economic actors, autonomously paying for third-party services using Stripe’s new machine payment protocol to plan a birthday party from a single prompt.

These workflows aren't just hypotheticals; they are deeply integrated into Stripe's engineering culture and point towards a future where software development is more accessible, and agents can transact on our behalf. Let's get into how they built it.

Workflow 1: From Slack Emoji to Shipped Code with Minions

One of the biggest challenges in a large organization is the friction between a great idea and actually getting it built. It's rarely about a lack of will; it's about coordination costs, technical access, and operational hurdles. The Stripe team tackled this by creating "Minions," AI agents designed to work within their existing developer environment to automate coding tasks.

Steve showed me how this dramatically lowers the activation energy for starting new work. Instead of needing to set up an environment, find the right files, and write boilerplate code, you can now trigger the whole process from a conversation.

Step-by-Step: Activating a Minion

Here’s the process Steve walked me through, which starts from a simple message in a Slack channel.

1. Start with a Prompt in Slack: An idea begins as a message. In Steve's example, he wants to improve the documentation for a new feature. He posts his idea as a regular message:

I have this cool idea for docs at stripe.com/payment/machine. This is our new machine to machine payment work, which we'll look at later in our call. Um, and I want to make sure the landing page really sticks and gives a good code. Example of how to get started quickly.

1. Trigger with an Emoji: To turn this idea into a task for an agent, all he has to do is add a specific emoji reaction, in this case, :create-minion-payserver:. The name payserver refers to the specific repository the work needs to be done in.
2. Automated Environment Provisioning: Once triggered, a bot confirms the Minion is being created. Crucially, it spins up a fully-configured, cloud-based development environment. This is built on years of investment by Stripe's developer productivity team. It checks out a new git branch, sets up the database, and even installs a VS Code server. This means you can run many agents in parallel without ever touching your local machine.

As Steve pointed out, "not only can I have one of these, but I could have many, many of these running in parallel in isolated environments, making isolated changes all at the same time."

1. The Agent Loop Begins: The Minion uses an open-source agent harness called Goose, which Stripe forked and adapted. The agent takes the original Slack message as its core prompt and begins a loop to solve the task. It uses Stripe’s internal tools to search the codebase, identify the correct files to modify, make the changes, run tests, and commit the code.
2. Pull Request for Human Review: Once the agent has completed the task, it creates a pull request. This is where the human comes back in. The 1,300 PRs per week are still reviewed by engineers, but their time is shifted from writing code to reviewing it. This process is made safe by Stripe's robust CI environment, extensive test coverage, and blue-green deployments that allow for quick rollbacks.

What’s brilliant here is how this system leverages existing investments. As we discussed in the episode, what's good for the human developer is good for the agent. Stripe's excellent documentation, tooling, and CI/CD a perfect foundation for their Minions to build upon, creating a virtuous cycle of productivity.

Workflow 2: AI Agents that Spend Money

In the second half of our conversation, Steve showed me something that truly feels like the future: an AI agent that can autonomously spend money to accomplish a task. This workflow explores the idea of agents as economic actors, capable of interacting with a new class of API-first businesses.

To demonstrate this, Steve tasked an agent powered by Claude Code with a fun but complex objective: plan a birthday party for his product manager, Jen.

Step-by-Step: Planning a Party for $5.47

The goal was to go from a simple prompt to a fully planned party, with the agent making real payments to real services along the way.

1. A Complex Prompt: Steve started with a detailed prompt in his terminal, telling the agent what it needed to do:

Research Jen Lee, who's my product manager, figure out what would be a good idea for her birthday, find a place to have the birthday, send invites to the birthday. And then, you know, we've burned all these tokens along the way, so we should probably donate to Stripe climate at the end to make up for all the energy consumption.

1. Pay-per-use Web Browsing: The agent first needed to research Jen’s interests. Instead of just scraping a page, it programmatically paid Browser Base for a single, ephemeral browser session. It wrote and executed Playwright code to navigate her personal website and discovered that she is a "matcha obsessed baker working on a cookbook." The cost for this was a fraction of a cent.
2. Paid Venue Search: Armed with this new information, the agent used Parallel AI to search for matcha-themed venues in New York City. It identified a suitable matcha cafe on Bowery.
3. Generating and Mailing an Invite: Next, the agent needed to send a physical invitation. It generated a PDF for the invite locally. Then, it programmatically paid a service called Postal Form to print and mail the physical invitation. This is a perfect example of an agent using a service for a task it physically cannot do itself.

1. Carbon Offset and Final Receipt: Finally, following the prompt’s instructions, the agent made a $1.65 contribution to Stripe Climate to offset the carbon footprint of its token usage. The agent then outputted a final receipt, detailing every transaction. The total cost to plan the party? $5.47.

This workflow is so powerful because it makes the economics of AI tasks tangible. We're already paying for tokens, but this puts that cost side-by-side with payments for real-world services. It opens the door for what Steve calls "ephemeral, API-first businesses" that are designed to be consumed by agents, not humans. These businesses wouldn't need a fancy UI or a login system, just a hyper-useful API that an agent can pay for on-demand.

The Future is Efficient and Automated

Talking with Steve left me incredibly energized. The Minions workflow shows how AI can be integrated into existing development practices to remove friction and amplify what engineers can do. It's a pragmatic approach that's already delivering massive value at Stripe.

The machine payments workflow is a glimpse into a new economy. When agents can transact, it unlocks a whole new category of businesses and applications. Suddenly, you can build complex chains of services to solve problems in a way that was never before possible. The lines between tokens, data, and dollars start to blur.

Both of these workflows underscore a key theme: AI's true power isn't just about executing tasks faster, but about fundamentally lowering the activation energy required to bring ideas to life. If you're an engineering leader, the big takeaway is to invest in your developer experience—what's good for your humans is great for your future agents. And for everyone else, start thinking about what you could build if your agent could not only code, but also had a credit card.

---

Brought to you by

Thank you to our amazing sponsors for supporting the How I AI podcast:

• Optimizely—Your AI agent orchestration platform for marketing and digital teams
• Rippling—Stop wasting time on admin tasks, build your startup faster

Episode Links

• YouTube: https://youtu.be/o5Mi5SYSDnY

Find Steve Kaliski on Twitter and learn more about Stripe's work at stripe.dev.