Back/How I AI
How I AI

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

Stripe Engineer Steve Kaliski shows us how their internal AI agents, or 'Minions,' turn Slack messages into code, shipping 1,300 PRs a week, and demonstrates how agents can autonomously spend money to plan a birthday party.

Claire Vo's profile picture

Claire Vo

March 24, 2026·7 min read
Episode outline

Stripe is landing roughly 1,300 pull requests a week that receive no human assistance until review. The surprising part is not the number. It is where those pull requests begin: a Google Doc, a ticket, a Slack conversation, or one emoji reaction.

Steve Kaliski says he cannot remember the last time he started work in a text editor. He still ends up there, but the first move is usually closer to the idea. Stripe's Minions turn that idea into an isolated development environment, an agent loop, tested code, and a pull request.

In this episode of How I AI, Steve shows the infrastructure behind that low-friction start. He also gives an agent money and lets it buy the services needed to plan a birthday party. The two demos share one idea: an agent becomes more useful when it can reach the real tools and economic systems required to finish the job.

The product breakthrough is lower activation energy

A good idea inside a large company can die before anybody writes code. The blocker may be access to the right technical expertise, uncertainty about who owns the area, or the ordinary coordination cost of getting several people aligned.

Minions collapse the distance between the conversation and the first implementation. Steve can see a request in Slack, add the create-minion reaction for the relevant repository, and watch Stripe's Devbox agent provision the work.

In the demo, he asks for a stronger code example on the landing page for Stripe's machine-to-machine payment work. The message becomes the prompt. A reaction selects the pay-servers repository. Seconds later, the system creates a branch and starts a development environment.

The full sequence is documented in the Slack-message-to-pull-request workflow. The interaction is simple because the infrastructure beneath it is not.

A Minion is a prompt inside a complete Stripe development environment

Stripe's codebase is too large and interconnected to run on one laptop. Long before Minions, the company invested in hosted development environments with the code, services, database configuration, Git settings, editor access, and CI tools an engineer needs.

A Slack message from the Devbox app details the completion of an automated agent run on `testing-tree-stone-field`, including its status, a summary of changes, and the generation of a pull request for a task filled with 'TODO comm...'. The wider context mentions a 'code example'.

A Minion provisions one of those environments with a prompt already inside it. Stripe uses a customized version of the open-source Goose harness for the agent loop. The loop can read internal documentation, use Stripe's code-search and MCP tools, inspect test data, change files, run checks, commit the result, and open a pull request.

Steve's system prompt is almost comically plain: "Implement this task completely," followed by the Slack message. The sophistication lives in the environment and the tools, not in a theatrical prompt.

That is why developer experience and agent experience reinforce each other. Clear paths for common API changes, strong internal docs, reliable environment setup, and fast tests help a new human engineer. The same assets keep an agent from exhausting its context window while wandering through Stripe's codebase.

Cloud environments make parallel work real

Local worktrees are useful until three or four agents make a high-end laptop sound like an airplane. A hosted environment removes that physical limit. Every Minion gets isolated code and services, so Steve can launch several changes at once without making them contend for one machine.

He can start a Minion from Slack on his phone during the commute. By the time he arrives, the agent may have found the relevant files, made the change, run the loop, and prepared a pull request. Steve can join halfway through instead of paying the full startup cost.

This is an infrastructure recommendation, not a demand that every company build Stripe's system. Teams serious about multi-agent engineering should ask whether agents can create a reproducible environment without a laptop staying awake, and whether ten simultaneous tasks are truly isolated.

Coding faster moves the bottleneck to review

Minions do not merge their own work because authorship does not change the standard. Whether Steve or Steve's robot wrote the text, the pull request goes through normal human review.

That review is credible because Stripe already has strong CI, test coverage, end-to-end synthetics, and rollout infrastructure. Blue-green deployments provide a rollback path. Without those systems, a reviewer staring at a plausible diff in a huge codebase would have little reason to trust it.

Steve expects attention to move as coding gets cheaper. Review may become the scarce resource. The company may need more good ideas, more distribution, or better prioritization. Saving authoring time does not eliminate product development; it reveals the next constraint.

Slack also makes the starting point available to people who do not identify as engineers. A product manager can write a brief. A designer can describe feedback. A customer-facing teammate can request a docs fix. Clicking an emoji is less intimidating than opening an editor, while the resulting code still enters the engineering review system.

Then the agent spent real money

The second demo moves from agents as internal workers to agents as economic actors. A model can already spend tokens, search the web, and call tools. Stripe's question is what happens when completing the prompt requires a paid third-party service.

Steve gives Claude Code a deliberately personal job: research his product manager Jen, choose a fitting New York birthday venue, send invitations, and offset the energy used by the run. The agent receives access to services that accept Stripe and Tempo's machine payment protocol.

The implementation pattern is in the agent-payments workflow. Instead of Steve creating accounts, choosing subscriptions, and entering a card, the agent buys the exact unit of service it needs.

Research

The agent pays Browserbase for one browser session, writes Playwright code, and opens Jen's site. It learns that she is a matcha-obsessed baker working on a cookbook. The session costs a fraction of a cent and ends when the research is complete.

A detailed AI prompt and API specifications for paid services being displayed in a code editor, demonstrating how an AI agent can access external tools for tasks like web search, content extraction, and sending physical mail.

Next it pays Parallel AI to search for suitable venues and selects a matcha cafe on Bowery. The service is not a permanent subscription in Steve's tool stack. It is a capability purchased for this task.

Act

The agent creates a PDF invitation and pays Postal Form to send physical mail. This is the boundary between what the model can do locally and what requires an outside operator. The robot can write the PDF. It cannot put an envelope into the postal system.

Account

The run finishes with an itemized transaction ledger. After using about 70,000 tokens, the agent contributes $1.65 to Stripe Climate, which the demo estimates offsets 4.4 kilograms of carbon. The complete birthday plan costs $5.47.

The receipt puts tokens and dollars beside each other. Both are resources spent to acquire intelligence, data, an operation, or an external service. We already accept that a coding prompt has an economic cost even when a subscription hides the penny counter. Machine payments make the rest of the task legible too.

The guardrails belong at the payment layer

Giving an agent a wallet is not the same as handing it an unrestricted corporate card. The system needs clear limits on total spend, approved services, payment credentials, data exposure, and which actions need confirmation.

The party is a playful demo, not proof that every consumer should plan events in a terminal. The product opportunity is a service designed for ephemeral agent use: one useful API, one paid interaction, no dashboard or long account setup required.

That changes the definition of a customer. Steve has already received detailed implementation feedback written by the agents that integrated Stripe's machine-payment docs. The humans asked their coding tools to read the docs, implement the feature, and then report what was confusing. The service provider increasingly has to hear from the agent directly.

A detailed AI agent's output displayed in a terminal, showcasing a 'Birthday Party Plan' with a 'Transaction Ledger' itemizing API calls, costs, estimated token consumption, and climate contributions.

What to copy from Stripe

The emoji is not the strategy. The strategy is a blessed path from idea to isolated environment to tested change to human review. Teams can build toward that path in layers:

  • Make development environments reproducible and available away from a single laptop
  • Document the common code changes and expose the tools agents need
  • Let work start where the idea already lives
  • Keep CI, human review, staged rollout, and rollback independent of who wrote the code
  • When agents spend money, expose the budget and ledger as part of the result

Minions are powerful because Stripe did not bolt a chat box onto a fragile development process. The company gave agents the same well-paved environment its engineers use, then lowered the activation energy until a Slack reaction could start the work.

Sponsors

Thanks for supporting How I AI

Optimizely

Your AI agent orchestration platform for marketing and digital teams

Rippling

Stop wasting time on admin tasks, build your startup faster

Build your next product with ChatPRD

Turn an idea into a PRD, user stories, and a plan.

Try ChatPRD free

Start shipping
better products.

Join 100,000+ product managers who use ChatPRD to write better docs, align teams faster, and build products users love.

Free to start
No credit card
SOC 2 certified
Enterprise ready