How I AIStripe's "Minions": How AI agents write 1,300 PRs weekly with 0 human coding
CHAPTERS
- 0:00 – 3:13
Stripe ships 1,300 AI-written PRs/week: what “Minions” are and why they matter
Steve Kaliski opens with the headline result: ~1,300 pull requests per week written by agents with humans only doing review. He frames Minions as a way to start work directly from where ideas live (Docs, Jira, Slack) and let agents carry tasks to a PR.
- •~1,300 PRs/week created without human coding (review only)
- •Work often starts in Google Docs/Jira/Slack—not the editor
- •A single emoji can trigger an end-to-end agent attempt (“one-shot”)
- •Humans jump in later for tweaks and review
- 3:13 – 5:53
Lowering activation energy beats faster execution in large orgs
Claire and Steve discuss how organizational friction slows good ideas, and why AI helps by reducing coordination and startup overhead. The big gain is making it effortless to begin, so momentum builds before a human even opens an IDE.
- •Big-company friction is often access/ops/silos—not malice
- •AI reduces coordination, communication, and execution costs
- •Key benefit: reduced activation energy to start work
- •Engineers can join mid-stream after agents build momentum
- 5:53 – 6:53
Minion architecture: hosted dev environments + tools + an agent loop
Steve explains the technical foundation: Stripe’s pre-existing hosted dev environments, CI, and internal tooling. A Minion provisions an env, seeds a prompt, and iterates using Stripe docs, CI, search, and test data to complete the task.
- •Stripe codebase is too large to run locally; cloud dev envs are core
- •Minion = provision env + seed prompt + run tool-using loop
- •Uses internal docs, CI, code search, and other Stripe systems
- •Goal is an end-to-end attempt that culminates in a PR
- 6:53 – 9:40
Demo: launching a Minion from Slack with an emoji reaction
In a Slack channel, Steve writes a natural-language request to improve a docs landing page, then triggers a Minion via an emoji reaction. The bot creates a branch, provisions a dev box, and starts the docs service so results can be previewed.
- •Slack is a primary “entry point” for initiating work
- •Emoji reaction triggers a specific repo-targeted Minion
- •Agent auto-creates a branch and kicks off services for preview
- •Provisioning step is fast because it builds on existing dev tooling
- 9:40 – 11:24
DX is agent experience: investing in developer tooling makes agents succeed
They emphasize a virtuous loop: what improves human developer experience also improves agent success rates. Blessed workflows, documentation, and standard pathways prevent agents from getting lost in huge codebases and reduce token/context costs.
- •“What’s good for developers is good for agents”
- •Blessed paths for common changes raise one-shot success rates
- •Good internal docs prevent agents from burning context window/tokens
- •DX investments become easier to justify when tied to AI velocity
- 11:24 – 13:51
Inside the agent run: minimalist system prompts, tool use, and “pair prompting”
As the Minion begins executing, Claire points out Stripe’s surprisingly simple system prompt style. Steve explains how Stripe augments prompting via bots, MCP servers, and collaborative “pair prompting” to construct better task context.
- •Simple system prompt + strong harness can outperform verbose prompting
- •Agents track to-dos and locate relevant files autonomously
- •Stripe uses MCP servers broadly to access internal systems
- •Prompt construction can be collaborative (humans + agents + data)
- 13:51 – 16:06
Why Stripe chose Goose as the agent harness (and how it fits with other tools)
Steve explains the decision to build Minions atop Goose: it provides a base loop/harness that Stripe can customize for its environment and workflows. Stripe still uses commercial tools (Claude Code, Cursor), but Minions provide the cohesive end-to-end internal workflow.
- •Goose provides a modifiable base harness/loop
- •Stripe forked/modified Goose to fit internal workflows
- •Minions are tightly coupled to Stripe’s dev environment and tooling
- •No single tool wins; Minions complement Claude Code/Cursor
- 16:06 – 17:15
Developer Productivity as an internal product: the team behind the platform
Claire probes organizational ownership, and Steve describes Stripe’s long-running developer productivity team. They treat internal engineering enablement like a product, spanning git workflows, editor configs, environments, and CI—now accelerated by AI use cases.
- •Dedicated developer productivity team exists long-term at Stripe
- •Owns git/versioning workflows, editor tooling, and dev environments
- •AI increases ROI/urgency of internal productivity investments
- •Minions ride atop years of platform/DX work
- 17:15 – 21:11
Cloud dev environments unlock true multi-threaded agent engineering
Claire and Steve argue local laptops are the bottleneck for parallel agent work. Cloud/virtual environments enable many isolated workstreams at once—agents can run in parallel, even triggered from a phone during commute time.
- •Local machines struggle with multiple worktrees/services
- •Cloud dev environments enable parallel, isolated changes
- •Agents can be kicked off asynchronously (even from mobile)
- •Analogy: modern dev requires parallelism like git branches did
- 21:11 – 22:06
What “one-shot” means in practice: one user prompt, many internal loops, one PR
They clarify the interaction model: the user provides one prompt, and the agent loops internally through tools, edits, and tests. The human re-enters at the end when there’s a finished PR to review, not during the execution steps.
- •One-shot UX: single prompt in, completed implementation out
- •Agent autonomously locates files, makes edits, runs checks/tests
- •Human involvement shifts to final tweaks + review
- •Creates momentum by making progress without constant supervision
- 22:06 – 23:50
Scaling review for 1,300 agent PRs/week: CI, test confidence, and shifting bottlenecks
Claire asks how Stripe reviews so many AI-authored PRs, and Steve points to rigorous CI, strong test coverage, and safe deployment practices (e.g., blue/green). He predicts that as coding gets cheaper, bottlenecks move to review, ideas, and distribution.
- •Review load is manageable partly because less time is spent coding
- •CI/test coverage is essential regardless of who wrote the code
- •Blue/green deployments and rollback maintain safety
- •As coding becomes “free,” review/ideation/distribution become bottlenecks
- 23:50 – 25:01
Minions beyond engineering: Slack as the universal interface for non-engineers
Steve explains why Slack makes Minions accessible to the whole company. Non-engineers can describe outcomes in plain language—product briefs, design feedback, doc changes—and trigger agents without needing local dev setup or deep technical familiarity.
- •Slack is shared across the entire org, lowering barriers to entry
- •Plain-English prompts let non-engineers initiate meaningful work
- •Emoji/tag-based triggering reduces intimidation vs. IDE workflows
- •Early signs of cross-functional adoption
- 25:01 – 35:07
Agents as economic actors: planning a birthday party with machine-to-machine payments
Steve demos a Claude Code workflow where the agent not only uses tools but pays for them programmatically. The agent purchases a BrowserBase session, uses Parallel AI for venue search, uses PostalForm to mail an invite, and offsets via Stripe Climate—ending with a cost breakdown.
- •Agents need the ability to spend money to accomplish real tasks
- •Pay-per-use access replaces human signup + credit card steps
- •Tools used: BrowserBase (browser session), Parallel AI (search), PostalForm (mail), Stripe Climate (offset)
- •Receipt-style output makes the economics of agentic work explicit
- 35:07 – 37:36
API-first, ephemeral businesses built for agents (and the future UX wrapper)
Claire and Steve extrapolate from the demo: today it’s “in the terminal,” but soon consumer-friendly wrappers will hide complexity. They predict businesses optimized for agent customers—minimal UI, a single strong API, and direct monetization via machine payments.
- •Terminal-driven workflows will be wrapped into friendlier UX soon
- •Emerging opportunity: businesses whose primary users are agents
- •Shift from dashboards/admin panels to “one great API” products
- •Agent-to-agent commerce enables new, ephemeral interaction models
- 37:36 – 41:55
Lightning round: personal AI workflows, prompting when agents fail, and where to follow
Steve shares a personal use case—building controlled, kid-friendly apps—highlighting “disposable software.” He describes failure-mode prompting: stay polite, ask the model to justify itself, and leave “breadcrumbs” via diffs/status to guide it; the episode closes with links to Stripe.dev and machine payments docs.
- •Personal workflow: fast prototyping “disposable” apps for parenting needs
- •Prompting strategy: politeness, request explanations/justifications
- •Guidance tactic: make partial progress then ask agent to inspect diffs/status
- •Resources: stripe.dev, docs.stripe.com/payments/machine, @stevekaliski
