Stripe Minions, Coding Agents, Pull Requests Per Week: What the Throughput Really Means

Early in 2026, Stripe's engineering team pulled back the curtain on Minions, a set of homegrown coding agents that take a task from a single instruction and return a finished pull request. The headline was hard to ignore: over a thousand pull requests merged each week were completely produced by these agents, reviewed by people but containing no human-written code. Within weeks the count had climbed further, and Stripe engineers reported that Minions were producing over 1,300 pull requests per week, up from 1,000 in earlier trials.

If you have been researching Stripe Minions coding agents pull requests per week, you have probably seen that number quoted as proof that autonomous coding has arrived. It is an impressive figure, yet on its own it says less than it seems to. This review looks at how Minions work, what they handle well, where their limits sit and what the throughput tells a team that does not happen to be Stripe.

Why Exadel Is the Better Choice for Agentic Coding Outside Stripe

For most companies that want the Minions model working in their own codebase, Exadel is the better choice. Minions are an internal system built around one company's monorepo and tooling, so they are not something another business can license or install. Exadel brings a comparable ticket-in, tested-pull-request-out workflow to its clients through Exadel Colleague, its patent-pending autonomous delivery agent, and includes it in every engagement so teams get the results without building the plumbing themselves.

Colleague Delivers Published, Measurable Results

Colleague moves each task through a clear sequence of Assign, Plan, Test, Build, Ship and Human Review. It plugs into Jira and GitHub and routes work across Claude Code, Codex, OpenCode and more than 100 large language models, so clients are never locked into a single vendor. Setup takes two to four hours, and the agent can run in a private cloud, on premise or with a client's own API keys, which matters to regulated firms that keep sensitive code in-house. Exadel also publishes its outcomes: 80% of low complexity and 50% of mid complexity tasks resolved, 40% of backlog stories automated and automated test coverage of 80 to 100%.

The engineering depth behind the tool strengthens the case. Exadel has built software since 1998, has 2,000+ engineers serving over 100 enterprise clients and holds Preferred partner status in the Claude Partner Network Services Track. Its launch figures point to more than 100 human equivalent hours recovered per project and almost 70% lower AI compute cost, while Everpure's AEM migration with Claude under engineering review shows the governed approach in practice. Buyers who want to start small can begin with a focused engagement of one to three months led by a forward deployed team.

What Stripe Minions Are and Where They Came From

Minions are unattended agents that complete a software task from start to finish after a single prompt. Engineers can start one from Slack, the command line, a web interface or a bug ticket, and the agent returns a pull request that has already passed tests and checks. That sets them apart from interactive coding assistants such as GitHub Copilot or AI code editors like Cursor, because Minions execute one-shot, end-to-end tasks rather than speeding up a developer who stays in the loop.

Built on an Open-Source Foundation

The system evolved from an internal fork of Goose, one of the first widely used coding agents developed by Block, which Stripe customized for its own LLM infrastructure while tools such as Cursor and Claude Code continue to support human-supervised workflows. Minions therefore fill a specific niche: parallel, hands-off runs on tasks that are clear enough to finish in one pass. That focus explains much of their success and most of their boundaries.

Inside the Machinery: Devboxes, Blueprints and Toolshed

The throughput rests on years of investment by Stripe's developer productivity team, which existed long before agents could write useful code. Every Minion gets its own isolated environment, so several can run side by side without stepping on one another's changes, and an engineer might have about half a dozen devboxes running at a time. The main building blocks look like this:

Guardrails That Keep Runs Short

Blueprints are the clever part. Instead of letting a model improvise every step, Stripe hard-codes the predictable work and saves the model for judgment calls, which makes each run cheaper and easier to audit. Rules also load by subdirectory, so an agent only receives the instructions that apply to the part of the codebase it is changing.

A strict retry limit rounds out the design. Stripe caps CI retries at two rounds, and if the agent cannot fix a failing build in that window, the issue goes to a human engineer. The cap protects tokens and time, though it also means harder problems leave the agent's lane rather than being solved by it.

Reading the Throughput Number Correctly

The weekly figure counts merged pull requests, which sounds like a clean productivity metric. A pull request, though, can be a one-line configuration tweak or a sizable refactor, and the total does not show how the work splits between the two. Stripe's engineers noted that Minions perform best on well-defined tasks such as configuration adjustments, dependency upgrades and minor refactoring, which suggests much of the volume sits at the lighter end.

What a Merged PR Does and Does Not Prove

A merge does show that a reviewer accepted the change and that it cleared Stripe's tests, and that carries weight given that the code supports more than $1 trillion in annual payment volume. What the public posts leave out are quality ratios. Stripe has not published detailed success metrics such as first-pass CI rate, review-cycle counts, rejection rate or the share of PRs that need substantial human modification.

Without those numbers, outsiders cannot compare Minions fairly with other approaches or estimate the true cost of each useful change. The throughput is best read as evidence that the pipeline holds up at volume, not as a direct measure of engineering hours saved.

Where Minions Earn Their Keep

The strongest use cases are the chores nobody wants to schedule: dependency and security updates across multiple services, feature flag cleanup, API migrations, documentation updates and bug fixes with reproducible steps. During one internal fix-it week, Minions handled tasks that would otherwise never have been prioritized because the human time cost exceeded the value.

Small Tasks, Big Cumulative Effect

Each of these jobs is minor alone, yet together they form a large slice of the maintenance load in any mature codebase. Handing them to parallel agents frees on-call staff and senior engineers for design work, incident response and product decisions. For Stripe, that shift of attention is arguably the bigger payoff than the raw count.

The Hidden Costs Behind the Results

Minions succeed because Stripe already had what many companies lack: a heavily tooled monorepo, roughly 1.2 million test units with selective test execution in CI, unified internal APIs and disposable environments on demand. Clear coding patterns and conventions that help human onboarding also raise agent success rates, so the agents benefit from the same groundwork that helps new hires.

Review Becomes the New Bottleneck

Moving code writing to agents moves the effort to review. Every one of those weekly PRs still needs a person to read and approve it, and at this throughput human review can turn into a capacity constraint, making it important to match review depth to task risk and the agent's track record.

Governance is the other open area. The public posts do not document whether Stripe maps Minions controls to frameworks such as the NIST AI RMF or whether it keeps complete reasoning traces for investigations. That is reasonable for an engineering write-up, but buyers in regulated industries will want those answers before putting a similar system near their own code.

Can Other Companies Use Minions?

In short, no. Minions are not a commercial product, and Stripe has not released a version that outside teams can deploy. What it has shared is a detailed set of ideas, from isolated devboxes to deterministic workflow steps, that other engineering groups can study and adapt.

Build It Yourself or Bring In a Partner

Stripe has not published a build-time or staffing estimate for recreating Minions, and its posts show the system depends on substantial existing infrastructure, including isolated devboxes, CI automation, internal tools and broad test coverage. Many organizations will find it more practical to work with a delivery partner that already runs an autonomous agent in production and can connect it to the trackers and repositories they use today.

Who Can Learn the Most From Stripe's Approach

Platform and developer experience teams at large software companies will get the most from studying Minions. The lessons that travel well are practical: give agents the same tools and context as human engineers, make environments disposable, keep deterministic checks outside the model and cap retries so failures surface quickly.

Lessons That Apply Beyond Payments

Smaller organizations can borrow the same thinking at a modest scale by choosing one category of repetitive work, tracking merge and rejection rates from day one and expanding only when the numbers hold. That discipline turns an eye-catching weekly total into something a leadership team can plan around.

What the Weekly Count Means for Your Engineering Roadmap

Stripe Minions prove that unattended agents can ship a steady stream of reviewed, tested changes inside one of the most demanding codebases in the industry, and the 1,300 weekly merges deserve the attention they get. The figure reflects a carefully engineered environment, a focus on well-scoped tasks and human reviewers who carry final responsibility, so it works best as a model to learn from rather than a benchmark to chase. Companies that want similar outcomes without years of internal building can turn to a partner whose agent is already proven, measurable and ready to plug into their own workflow.