Stripe’s Autonomous Coding Agents Generate Over 1,300 PRs a Week

Global fintech giant Stripe says more than 1,300 pull requests merged every week across its engineering organisation are now generated end-to-end by its internal unattended coding agents, called “minions”. 

Minions are Stripe’s internal unattended coding agents that handle tasks from fixing flaky tests to implementing features — start to finish, with no human in the loop. The company detailed its efforts and reasonings as to why it built its own coding agents in a two part blog post.

A typical minion run begins with a Slack message and ends with a pull request that has already passed CI. Engineers never touch the keyboard between those two events. 

Each minion runs on a “devbox” — a standardised AWS EC2 instance pre-loaded with Stripe's source code and services, provisioned in under 10 seconds from a warm pool. The isolated environment means minions run with full shell permissions without needing confirmation prompts, since any mistakes are contained to a single machine.

The agent harness is a fork of Block's open-source coding agent Goose, heavily modified for unattended operation. Central to Stripe's implementation is a construct called blueprints — orchestration flows that mix deterministic code nodes with open-ended agent loops. 

The company describes the design principle as “putting LLMs into contained boxes compounds into system-wide reliability upside.” 

A node labelled “Run configured linters” executes fixed code; a node labeled “Implement task” gives the model latitude to reason and act. The result is a state machine that guarantees certain steps happen regardless of model behaviour.

For dynamic context, minions connect to Toolshed, Stripe's centralised MCP server. Stripe describes it as providing “a common language for all agents at Stripe, not just minions,” with nearly 500 tools spanning internal systems and third-party SaaS platforms. 

Each minion receives a curated subset relevant to its task. Feedback loops run locally before CI is ever invoked, shifting lint checks and test heuristics left. The standard blueprint allows at most two CI rounds before returning the branch to a human.

Stripe built its own minions in-house rather than relying solely on third-party agents due to the scale and complexity of its internal systems. 

The company operates hundreds of millions of lines of mostly Ruby code, relies heavily on proprietary libraries, and runs infrastructure that processes more than $1 trillion in annual payment volume. These conditions make it difficult for general-purpose agents to reliably follow internal conventions and compliance constraints.

Minions are tightly integrated with Stripe’s existing developer infrastructure rather than treated as a standalone AI tool. “What’s good for humans is good for agents, and building on this infrastructural primitive paid dividends as a natural home for LLM agents,” said Stripe.