Stripe Deploys Autonomous Coding Agents
Stripe has deployed autonomous AI coding agents that are now shipping 1,300 pull requests per week. These agents are being used to automate routine data engineering and ML tasks, and are beginning to contribute to pricing, claims, and risk systems. The move signals a broader enterprise shift toward using agentic AI for rapid prototyping, compliance checks, and the continuous improvement of software.
- Stripe's internal AI coding agents, known as "Minions," are built on a fork of Block's open-source coding agent, "goose". These agents operate in isolated cloud-based developer environments called "devboxes," which are the same environments used by human engineers at Stripe. This setup allows for safe, parallel execution of multiple AI agents without interfering with production systems or each other. - The agents are triggered by engineers via Slack messages, a command-line interface, or other internal applications like ticketing systems. They can access the full context of a discussion thread, including links to documentation and ticket details, to inform their work. All pull requests generated by the agents still require human review before being merged. - To interact with Stripe's internal systems, the agents use a centralized server called "Toolshed," which hosts nearly 500 tools built on the Model Context Protocol (MCP). This allows the agents to access a curated set of internal documentation, code intelligence from Sourcegraph, and other necessary data. - Other major tech companies are also heavily investing in AI-driven software development, with Google reporting that over 30% of its new code is AI-generated, and Meta's Mark Zuckerberg stating a goal for AI to handle half of the company's software development. - The AI agents' workflow combines creative steps performed by a large language model with deterministic, hardcoded steps for tasks like running linters and making git commits. If a test fails after an agent pushes code, the error is fed back to the agent for a fix, but it is limited to a maximum of two attempts before escalating to a human engineer. - Stripe's codebase consists of hundreds of millions of lines of code, primarily in a custom Ruby framework with Sorbet typing, a stack that is not commonly found in the training data of large language models. This necessitates the use of specialized internal tooling and context to make the agents effective. - In addition to coding agents, the insurance industry is exploring agentic AI for tasks like claims triage, fraud flagging, and generating personalized policy offers. This reflects a broader shift from AI that suggests actions to AI that can execute complete, complex processes autonomously within set guardrails. - To facilitate an economy where AI agents can transact, Stripe has introduced a machine payments feature and co-developed the Agentic Commerce Protocol (ACP) with OpenAI. This allows developers to charge autonomous agents for API access and other services using microtransactions.
