Quality assurance · Production

Stripe Minions: Over 1,300 AI-produced pull requests merged weekly via unattended agentic coding on devboxes

The problem

Scaling unattended agentic coding at Stripe required a parallelizable, predictable, and isolated environment that was fundamentally difficult to achieve with local machines, and off-the-shelf coding agents were optimized for human-supervised use rather than fully unattended runs.

Workflow diagram · grounded in source
1
Devbox provisioned hot and ready
trigger
“we aim for it to be ready within 10 seconds. To achieve this "hot and ready" standard, we proactively provision and warm up a pool of devboxes so they are ready when a developer wants them”
2
Rule-file context gathered
integration
“we almost exclusively give minions context from files that are scoped to specific subdirectories or file patterns, automatically attached as the agent traverses the filesystem”
3
Dynamic context via MCP tools
integration
“to fully hydrate user requests, minions need to retrieve information such as internal documentation, ticket details, build statuses, code intelligence, and more”
4
Blueprint agent implements task
ai_action
“there are agent-like nodes with labels such as "Implement task" or "Fix CI failures." Those agent nodes are given wide latitude to make their own decisions based on input”
5
Deterministic lint node runs
validation
“We run a subset of linters as a deterministic node within the agent devloop blueprint, and loop on that lint node locally before pushing an agent's branch, so that the branch has a fair shot at passing CI the first time around”
6
Push and CI run with autofixes
output
“After a minion pushes a change, we run CI and auto-apply any autofixes for failing tests”
7
CI failure feedback and retry
feedback_loop
“If there are failures with no autofix, we send the failure back to a blueprint agent node and give the minion one more chance to fix the failing test locally. After the second push and CI run, we send the branch back to its human operato…”
8
Human review of PR
human_review
“completely minion-produced, human-reviewed, but containing no human-written code”
Reported outcome

Over 1,300 Stripe pull requests are merged each week that are completely minion-produced and human-reviewed but contain no human-written code.

Reported metrics
minion-produced PRs merged per weekover 1,300
Devbox ready timewithin 10 seconds
MCP tools in Toolshednearly 500
Stripe automated test suite sizeover three million
Reported stack
gooseToolshedMCPBazelCursorClaude CodeAWS EC2
Source
https://stripe.dev/blog/minions-stripes-one-shot-end-to-end-coding-agents-part-2
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Over 1,300 Stripe pull requests are merged each week that are completely minion-produced and human-reviewed but contain no human-written code.

What tools did this team use?

goose, Toolshed, MCP, Bazel, Cursor, Claude Code, AWS EC2.

What results were reported?

minion-produced PRs merged per week: over 1,300; Devbox ready time: within 10 seconds; MCP tools in Toolshed: nearly 500; Stripe automated test suite size: over three million (source-reported, not independently verified).

How is this quality assurance AI workflow structured?

Devbox provisioned hot and ready → Rule-file context gathered → Dynamic context via MCP tools → Blueprint agent implements task → Deterministic lint node runs → Push and CI run with autofixes → CI failure feedback and retry → Human review of PR.

Related quality assurance cases