Lore

AI coding agents

Stripe's "Minions": How AI agents write 1,300 PRs weekly with 0 human coding

Stripe's Steve Klaliski argues that Stripe now ships ~1,300 PRs a week with no human authoring beyond review because the company built "minions" — Slack-triggered, cloud-hosted coding agents layered on years of developer-experience investment (docs, CI, hosted dev environments) — and that the same infrastructure is now being extended so agents can act as economic actors, spending real money via a Machine Payment Protocol to complete tasks, which he argues previews a coming "agent economy."

How I AI · 2026-03-25 · English

Key ideas

  1. Most friction between a good idea and shipping it in a large org is functional or operational, not malicious; Stripe's "minion" pattern (click an emoji in Slack/Jira/a doc) removes that friction by letting an agent one-shot resolve the prompt using all internal tools, so work now starts in Slack or a doc rather than a text editor.

  2. Confidence in shipped code comes from CI, test coverage, synthetics, and blue-green rollback — infrastructure that is agnostic to whether a human or an agent wrote the code, so Stripe's existing pre-AI investment in developer tooling is what makes agent-authored PRs safe to land at volume.

  3. "What's good for the developer is good for the agent": good documentation, a well-trodden "blessed path" for common tasks, and MCP-connected internal tools raise both human and agent success rates, creating a virtuous loop that Steve says can be used to justify funding long-deferred developer-experience work.

  4. Local hardware (laptops, a handful of work trees) is a hard ceiling on how many agents can run in parallel; Steve argues cloud/virtual dev environments are the necessary infrastructure to unlock multi-threaded agentic engineering, and urges CTOs/VPs of engineering to invest there in the next year.

  5. As AI drives coordination, execution, and communication costs toward zero and code-writing becomes cheap, Steve argues the bottleneck in the development process shifts elsewhere — to code review, or to generating/distributing enough good ideas — rather than disappearing.

  6. Because triggering a minion only requires a plain-text prompt in Slack (not a text editor), Stripe sees growing non-engineer usage (PMs, designers) writing prompts like product briefs or design feedback, extending the same low-activation-energy pattern beyond engineering.

  7. Stripe frames agents as economic actors that must be able to spend money, not just tokens, to fulfill a prompt; the birthday-party demo (using a Machine Payment Protocol co-designed with Tempo) shows an agent paying Browserbase, Parallel AI, and Postal Form on the fly instead of a human provisioning access/credentials in advance.

  8. Steve argues token costs and real-currency costs are converging conceptually — spending tokens on an LLM and spending dollars on a third-party service are becoming functionally the same activity — implying future businesses may skip dashboards/landing pages and instead monetize a single API directly to agent customers.

  9. Effective agent collaboration technique mirrors directive, corrective human guidance rather than either full manual control or blind trust: ask the agent to explain/justify itself, leave "breadcrumbs" (git diffs) for it to follow, and codify repeatable approaches into reusable skills/prompts.

  10. Minions — Stripe's internal name for Slack-triggered, cloud-hosted coding-agent tasks that attempt to one-shot resolve a prompt using all of Stripe's internal tools. Apply: Click an emoji reaction on a Slack message, Jira ticket, or Google doc naming the target repo (e.g. "create minion pay server") to spin up a provisioned dev environment and let the agent implement the task end-to-end into a reviewable PR.

  11. Goose — An open-source agent harness originally built by Block, forked internally by Stripe as the base tool-calling loop underneath minions. Apply: Boot an instance of Goose with a freeform task prompt and let it loop through tool calls (search, edit, test, commit) until it returns a single final "successful implementation" response.

  12. MCP (Model Context Protocol) — A protocol Stripe used to expose nearly all of its internal data sources and tools to agents. Apply: Stand up an MCP server for each internal system (docs, code search, test data, CI) so agents (minions or Claude Code) can call them directly instead of relying only on model knowledge.

  13. Blue-green deployment — A deployment strategy that runs two environments so a release can be rolled back quickly if it fails. Apply: Route production traffic to the new ("green") environment only after validation, keeping the old ("blue") one live for instant rollback regardless of whether the change was written by a human or an agent.

  14. Pair prompting — The replacement for pair programming: collaboratively crafting a prompt with another engineer, a data source, or another agent before execution. Apply: Before triggering a minion, work with a colleague or a prompt-writing bot to refine the task description so the agent has enough context to one-shot the work.

  15. Blessed path / documented workflow — A well-documented, standardized procedure for a common engineering activity (e.g. adding a new API field) performed hundreds or thousands of times a year at Stripe. Apply: Invest in writing and maintaining canonical documentation for repeated task types so both human engineers and agents can follow the same proven steps and succeed more reliably.

  16. Minimal system prompt philosophy — A deliberately terse system prompt ("implement this task completely:" plus freeform content) that relies on the harness's quality rather than elaborate prompt engineering. Apply: Avoid over-architecting prompts; trust a well-built harness (docs, tools, blessed paths) to extract a successful outcome from a loosely specified task.

  17. Prompt-writing bots — Internal Stripe bots that search the codebase, PRs, and Google docs to help author the task prompt before an agent executes it. Apply: Run a prompt-drafting bot against the relevant context sources first, then hand the refined prompt to the minion/agent for execution.

  18. Self-maintained agent to-do list — A codified focus mechanism where the agent tracks its own task list during execution. Apply: Let the agent break a task into a visible to-do list it updates as it works, so progress and remaining steps are inspectable mid-run.

  19. Machine Payment Protocol — A protocol co-designed by Stripe and Tempo that lets an AI agent make real payments to third-party services as part of fulfilling a prompt. Apply: Equip an agent with payment credentials/capacity so it can pay-per-use for services (browser sessions, search, mailing) on demand instead of a human pre-provisioning access.

  20. Browserbase — A pay-per-session hosted browser automation service with no prior signup required. Apply: Have the agent pay for a new browser session on demand and drive it with locally-generated Playwright code to interact with live websites.

  21. Playwright — A browser automation library the agent used to write code locally that connects to a remote Browserbase session. Apply: Generate Playwright scripts on the fly to control a paid browser session for tasks like navigating a website to gather information.

  22. Parallel AI — An AI-driven web search service used by the agent to find real-world options (e.g. NYC venues). Apply: Call Parallel AI as a paid search tool to surface candidate results matching criteria extracted earlier in the task.

  23. Postal Form — A service that takes a PDF and physically mails it. Apply: Generate the document locally (e.g. an invite PDF) then pay Postal Form to handle the physical mailing step an agent cannot perform itself.

  24. Agent receipt — An itemized record showing every service an agent transacted with during a task and its cost. Apply: Surface a receipt after an agent run to make the economics of an agentic task (which services were paid, how much) transparent to the user.

  25. Stripe Climate — Stripe's carbon-offset purchase mechanism. Apply: Have the agent calculate the carbon footprint of the tokens/compute used for a task and purchase an offset (e.g. $1.65 for 4.4kg) as a final step.

  26. Usage-based / LLM token billing — A Stripe billing model (including a newly announced beta) that supports charging based on usage, extended to LLM/token consumption. Apply: Use Stripe's usage-based billing infrastructure to charge customers according to tokens consumed or agent actions taken, rather than flat subscriptions.

  27. VS Code server (web) — A browser-accessible instance of VS Code automatically configured inside each minion's provisioned dev environment. Apply: Pop into the VS Code web instance tied to a minion's environment to manually inspect or tweak the agent's changes before merging.

  28. Explain/justify-yourself prompting — A technique of asking the AI to explain or justify its output when it isn't following instructions correctly. Apply: When an agent isn't one-shotting a task, prompt it to explain its reasoning to surface and correct the misunderstanding.

  29. Breadcrumbing via git status/diff — A guidance technique where the human starts work in the right direction and then points the agent at the resulting git diff instead of writing full explicit instructions. Apply: Make a small manual edit toward the correct solution, then ask the agent to review the git status/diff as directional context for completing the rest.

  30. Skill/prompt capture — Codifying a one-off-but-repeatable task approach into a reusable skill or saved prompt. Apply: After solving a non-recurring-but-likely-to-recur task, save the successful approach as a skill/prompt to inject back into future agent runs.

Insights

The minion pattern decouples starting work from being present: Steve can trigger a task from his phone on the subway and later "hop in halfway through," riding what he calls "generative momentum" rather than experiencing coding as synchronous labor.

Prompt-writing has itself become an engineered, semi-automated discipline at Stripe — bots search the codebase, PRs, and Google docs to help author the prompt before the agent ever runs, and engineers describe this as "pair prompting" rather than pair programming.

A real product-development feedback loop compressed from days to about 30 seconds: an engineer had Claude/Codex read Stripe's docs, implement the machine-payment integration, and have the agent itself send Steve a two-page feedback report — happening 4-5 times in a week, which Steve says gave him a physically "jarring" sense that agents are becoming literal users of Stripe's product.

Steve explicitly treats the DX/agent-experience flywheel as a political lever: attaching developer-experience roadmap asks to AI initiatives is presented as a tactic for getting them resourced.

The itemized "agent receipt" from the birthday-party demo (Browserbase session, Parallel AI search, Postal Form mailing, Stripe Climate offset) is used to make the argument tangible that agentic work already has a real, itemizable dollar cost even before any human-facing product exists around it.

Steve's admission that he deliberately stays polite to AI despite knowing rudeness sometimes yields better results — because he doesn't want it "on the record" that he was mean — surfaces an unresolved, only half-joking anxiety about how humans should relate to increasingly agentic systems.

Mid-interview, Steve notices unprompted that he is raising young children at the same moment he is building AI agents, and draws an explicit but uncertain parallel between techniques for guiding each (explaining/justifying, breadcrumbing, skills) without claiming one causes the other.

«At Stripe, we're landing about 1,300 PRs that have no human assistance besides review per week.»

— 00:00

«Whether the text has been written by Steve or the text has been written by Steve's robot, you still want that CI environment that's providing confidence that the code that's being changed is safe.»

— 00:35

«No matter how juiced these laptops are, you get three or four work trees in and like it starts to sound like an airplane taking off, it's no good.»

— 00:53

«So good docs for developers are uh equally important for the agent.»

— 11:11

«we don't pair program anymore but we you know pair prompt»

— 12:51

«It's like there's a there's a whole business around just the computer not falling asleep.»

— 19:43

«we have to equip our agents with the capacity to spend so they can not only consume tokens, but so they can also pay services as part of achieving the prompt.»

— 26:20

«You know you're doing something wrong if I have to load environmental variables to celebrate someone's birthday.»

— 29:17

«It's just a bunch of binary. No one's going to come to the party.»

— 30:19

«the token and the currency that backs it like they feel closer than ever»

— 34:19

«I planned a birthday party for $547. That doesn't seem too bad.»

— 35:03

«Agents all the way down.»

— 36:36

«There's something really weird about raising kids at the exact same time that the robot emerges.»

— 40:33

«It's really interesting to be raising kids and literally writing like soul.md files into my agents.»

— 40:47

Reception

Reactions are mixed, with genuine technical critiques and confusion alongside some enthusiastic praise and practical troubleshooting discussion.

This is a promotional-leaning but content-rich founder/engineer interview: it's most credible as a detailed map of Stripe's internal agentic-engineering stack (minions, Goose, MCP, blessed paths) and least verifiable in its headline PR-volume and cost figures and in its speculative claims about a coming machine-payment/agent economy.

41:56

↳ How I AI · YouTube

Watch original