Stripe's 'minions' ship PRs
Stripe’s internal AI coding agents — nicknamed “minions” — are now shipping roughly 1,300 pull requests per week, automating code changes and platform upkeep at scale. That shows AI agents can move from novelty to core dev‑platform productivity tools, and raises questions about orchestration, code review, and security for production pipelines. (lennysnewsletter.com)
Invocation for a minion run can start from Slack emoji reactions, a CLI command, or a web/internal UI tied to tickets and feature requests, with engineers often triggering agents directly from thread messages. (lennysnewsletter.com) Minions execute inside Stripe “devboxes”—AWS EC2 developer environments that Stripe pre-warms from a pool so a ready box can be handed to a run in roughly 10 seconds, including cloned repos and warmed Bazel/typecheck caches. (stripe.dev) Context and capabilities come from a centralized Model Context Protocol server Stripe calls Toolshed, which exposes more than 400 MCP tools (docs, tickets, build status, Sourcegraph, and SaaS integrations) that agents can call as a curated subset per task. (stripe.dev) The minion orchestration uses a fork of Block’s open-source Goose agent for the core loop and runs “blueprints” that interleave deterministic workflow nodes (git ops, linters, tests) with agentic nodes that perform implementation or fixes. (stripe.dev) Stripe shifts feedback left by running local lint and heuristic tests before CI, and its standard blueprint limits automated retries to at most two CI rounds before handing the branch back to a human reviewer. (stripe.dev) (analyticsindiamag.com) Operational constraints motivated an in‑house build: Stripe’s backend is predominantly Ruby with Sorbet, its repos amount to “hundreds of millions” of lines, and the production systems process well over $1 trillion in payment volume annually—factors Stripe cites as reasons to tightly integrate agents with existing developer tools. (stripe.dev) Engineers frequently run multiple isolated minion runs in parallel to parallelize small tasks (for example during on‑call rotations), with minion runs designed to use the same linters, CI pipelines, and rule files as human engineers to keep generated changes auditable and reviewable. (stripe.dev)
