Stripe's Minions case starts years before the agent. The company already had standard development computers, fast checks, a large test system, internal tools, and strong review habits. Minions turned that existing factory floor into an unattended workflow.

What are Stripe Minions?

A Minion is Stripe's internal background coding agent. A person delegates a task, often from Slack. The system gathers context, changes code in an isolated development computer, runs checks, responds to a bounded amount of automated build-and-test feedback, and returns a branch for inspection.

Stripe calls this “one-shot” because the employee can make one delegation and leave the run unattended. It does not mean one model call, autonomous merging, or delivery without review. The initiator inspects the result, then another Stripe engineer reviews the pull request.

How does a request become a reviewed pull request?

Stripe places Minion entry points where work and context already live: Slack, internal documentation, command-line and web tools, feature flags, and ticketing systems. A task receives a prewarmed AWS development computer based on the same standard environment humans use. Stripe reports that it becomes ready in roughly ten seconds.

The workflow is defined as a Blueprint. Blueprints alternate between two kinds of step:

  • Agent steps handle work that needs interpretation, such as understanding a request or changing code.
  • Deterministic steps perform known operations, such as preparing context, checking formatting and code rules, saving source-control changes, or starting continuous integration (CI), the automated build-and-test system.

This separation is a central software-factory pattern. Use model judgment where the path is uncertain. Use ordinary code where the rule is known and repeatable.

How does Minions get company context?

Stripe keeps guidance close to the code it describes. Directory and file rules supply only the instructions relevant to the current work. A deterministic retrieval step gathers useful documents and links before implementation begins.

Minions can also use Toolshed, Stripe's central internal Model Context Protocol (MCP) server. Stripe reported nearly 500 internal tools by February 19, 2026. Each Minion receives a curated subset instead of the entire catalog. This reduces clutter and limits what a run can access.

The development environments exclude real user data, production services, and arbitrary network access. That boundary lets the agent work with realistic code and tests while reducing the cost of a mistake.

What happens when tests fail?

Fast local code checks and known fixes run before CI. Stripe's account says the CI system selects applicable tests from a corpus of more than three million tests. That does not mean every test runs for every Minion.

The repair loop is deliberately finite. CI can apply known fixes, then the agent gets one repair attempt and a second CI run. After that, the branch returns to the person. A bounded loop prevents the agent from spending indefinitely on a weak approach.

What scale has Stripe disclosed?

Stripe reported more than 1,000 merged Minion-produced pull requests per week on February 9, 2026, and more than 1,300 ten days later. In an August interview, Stripe president Will Gaybrick said roughly 7,000 pull requests had come from Minions in the preceding week, about 30 percent of that week's total.

The later wording matters. “Came from” does not establish that all 7,000 were merged. None of these volume figures reveals completion rate, abandoned runs, review time, defects, cost, or a controlled productivity comparison.

Is Minions open source?

Minions began as an internal fork of Goose, the open-source agent that originated at Block. Goose remains public. Stripe's fork, Blueprints, Toolshed, and internal integrations are not public in the cited evidence.

What can a product team copy?

Inventory the factory floor before choosing a model. A strong background-agent pilot needs a standard environment, fast local checks, selective CI, scoped instructions, curated tools, safe data boundaries, a finite repair budget, and two clear human decisions: accept the proposed work and review the change.

Stripe's case suggests that background autonomy is often an infrastructure feature. The agent can leave the foreground because the environment, context, checks, and handoff are already dependable.