Inside Stripe's software factory
One-shot end-to-end coding agents on a forked Goose, gated by a hard two-round CI cap and a shared 500-tool internal MCP server.
Stripe at a glance
The company
| Sector | "A technology company focused on improving the conditions for economic growth and prosperity" 7 |
|---|---|
| Employees | Not disclosed as a discrete, source-confirmed count |
| Customers | 5m+ businesses powered directly or via platforms 8 |
| Revenue | $6.8bn 2025 revenue (+33%); $1.9tn total payment volume (+34%); $159bn valuation (Feb 2026 tender offer) 8 |
No first-party source discloses a company-wide or engineering-specific headcount. Third-party aggregators put total staff in the high thousands and disagree with each other on the number and on whether it's growing or shrinking, so the row is left unstated rather than guessed.
The engineering organisation
| Engineering size | Not disclosed. The Leverage team built and owns Minions, Toolshed, and Blueprints, but no headcount for that team appears in any source 1,2 |
|---|---|
| AI tooling since | Undated internally; the earliest dated public account is February 2026 (Minions Part 1) 1 |
| Tool policy | Built in-house, on a fork of Block's open-source Goose, rather than bought 1 |
| Agents in the SDLC | One-shot end-to-end coding agent (Minions); a shared internal MCP tool server (Toolshed) used by other agent systems beyond Minions 1,2 |
| Reported adoption | 1,300+ PRs merged per week as of February 2026, up from 1,000 ten days earlier 2 |
| Published metrics | PR volume per week, Toolshed tool count, test-suite size, CI round cap 1,2 |
Stripe's own Ruby monorepo is given as "on the order of 50 million lines" in one engineering post 6 and as 30 million lines in a separate conference talk by the Minions author 5; no source reconciles the two.
People
What they set out to solve
Stripe's public account frames the problem less as "generate code faster" and more as removing friction from the moment an idea occurs to the moment an agent starts acting on it. "Good developer experience for humans creates better outcomes for AI agents," is how Steve Kaliski, the engineer who leads the Minions effort, put it, framing the work as lowering the activation energy between having a task and getting an agent to run it 3.
The Leverage team, the internal group that builds developer-productivity tooling, describes the same problem from the infrastructure side. Once an agent can produce a complete pull request in one shot, iteration cost becomes the real bottleneck, so the system caps CI at two rounds per run and pushes local lint under five seconds, specifically to cut CI churn before a change ever reaches the shared build queue 1. A narrower, third problem sat one layer down: rather than let every internal system and SaaS integration expose its own bespoke agent hooks, Stripe built one shared front door, an internal MCP server called Toolshed, that any agent, not only Minions, reaches through 1,2.
Maturity level
Assigned by testing published facts against fixed criteria, and capped by what the sources state.
How Stripe's software factory works
Minions are one-shot, end-to-end coding agents built on a fork of Block's open-source Goose, the only traceable case in this corpus of one studied company building directly on another's agent code 1. A minion run starts wherever the work already lives: tagged in a Slack thread (including by emoji reaction), typed into a CLI or web UI, filed through the docs platform, the feature-flag system, or a ticketing UI, or triggered automatically when CI detects a flaky test 1,3. It ends with a CI-passing pull request and zero human-written code inside it, though every merge still requires a human reviewer 1,3,4.
Getting there runs through a hard-capped loop: local lint finishes in under five seconds before a branch is even pushed, specifically to keep noise out of CI; a first full CI run applies autofixes automatically; any tests still failing without an autofix go back to the agent for exactly one local attempt; a second and final CI run follows; then a human reviews the result 1. Stripe's own description of the loop draws a clean line: minions do the writing, humans do the merging, and the line doesn't move 1,3,4.
Toolshed is the tool layer underneath all of it: one internal MCP server that had 400 tools at launch and grew to roughly 500 within ten days, spanning internal systems and external SaaS platforms, shared with agent systems beyond Minions rather than built as a Minions-only integration 1,2. Blueprints sit above the tools, an orchestration primitive that interleaves deterministic workflow steps with agentic decision points, so a run isn't loose across the whole task, it's loose at specific, named junctions 1,4. Safety is handled environmentally rather than through heavy output moderation: minions run on the same devboxes engineers use day to day, pre-warmed in about ten seconds, but scoped to Stripe's QA environment with no access to real user data, and restricted to a curated default toolset 1,2.
A separate, unrelated engineering post about test-infrastructure gives a glimpse of the pressure this system creates elsewhere: Stripe's Ruby monorepo runs roughly 1.2 million test units across about 100,000 files, and a purely deterministic test-selection system built by a different team names "AI adoption among our engineering teams accelerating commit rates" as part of what made brute-force test execution unsustainable, independent, incidental confirmation that AI-driven commit volume is a real and felt load on Stripe's CI, from a team with no stake in the Minions story 6.
Operating the software factory: teams, roles, and skills
Minions, Toolshed, and Blueprints were built and are owned by the Leverage team, Stripe's internal developer-productivity group, with Alistair Gray publishing both parts of the system's public account 1,2. Steve Kaliski, described in his own podcast interview as the engineer who leads the Minions effort, is the source for how work actually enters the system, Slack threads, emoji reactions, dedicated team and personal channels 3. Cameron Bernhardt, an engineering manager, independently corroborates the weekly PR figure and the human-review requirement to trade press, without being part of the Leverage team itself 4. No source discloses how many people work on Minions, Toolshed, or Blueprints directly.
The clearest claim about who else uses the system is also the thinnest: Kaliski says a product manager can write a brief, a designer can describe feedback, and a customer-facing teammate can request a docs fix, and a minion will act on it 3. It's one line from one interview, with no corroboration or figure attached anywhere else in the record.
Implementation journey
- February 2026Minions Part 1 publishes, the first dated public account of an agent producing complete, CI-passing pull requests without human-written code at Stripe 1.
- February 2026Part 2 follows within ten days, reporting growth to 1,300+ PRs merged per week (from 1,000) and Toolshed's tool count climbing toward 500 2.
- March 2026
- April 2026A separate, AI-unrelated engineering post about test-selection infrastructure cites accelerating AI-driven commit rates as part of its own motivating pressure 6.
- May 2026Alistair Gray presents Minions at a live conference, citing a 30-million-line codebase figure and describing review patterns built around risk profiles and code ownership 5.
Measuring the software factory
The most rigorously sourced figure is PR volume week over week, 1,000 to 1,300+ within ten days, reported first-party and independently corroborated by an engineering manager speaking to trade press 1,2,4. Toolshed's tool count is reported on the same cadence, 400 growing to roughly 500 1,2.
Two other figures don't hold up as cleanly. Stripe's own test-suite size is given as "over three million" tests in the Minions series 2, while an unrelated testing-infrastructure post from a different team puts it at roughly 1.2 million test units across about 100,000 files 6; nothing in either source states whether they're counting the same unit. And the codebase itself is given as "on the order of 50 million lines" of Ruby in that same testing post and as 30 million lines in a later conference talk by the Minions author 5,6, with no source reconciling the two.
Non-engineer adoption and the machine-to-machine payment demonstration are both real claims in the record, but both arrive as single, unquantified statements from one interview rather than measured figures 3.
Lessons learned
Iteration is deliberately rationed, not just optimized: exactly two CI rounds per run, a design choice Stripe describes as being "specifically to cut CI churn" 1. Read next to the unrelated test-infrastructure post's own admission that accelerating AI-driven commit rates were already straining CI capacity elsewhere in the company 6, the cap looks less like a judgment about agent output quality and more like a scaling decision made under real infrastructure pressure.
A "40% of daily code is AI-generated" figure circulates on third-party listicle and marketing sites that blend industry-wide statistics with Stripe mentions; it could not be traced to any Stripe source. A claimed "$2,400/month saved per developer" traces to nothing first-party either. Neither should be treated as a Stripe metric. Stripe's own annual letter discusses agentic commerce and machine-to-machine payments at length, but that's product strategy for customers' AI agents, not a description of Stripe's internal engineering practice, and the two are easy to conflate because they use similar vocabulary 8.
What's next for Stripe's software factory
Stripe hasn't published a dated roadmap item. The closest the record comes is a secondary summary of Alistair Gray's May 2026 conference talk, which gestures at cloud-based development boxes, browser-based verification, and "command-center" workflows for future Minion iterations 5. That's a session-listing summary rather than a verified primary transcript, so it should be read as directional, not as a stated commitment.
