Stripe Minions | Ry Walker Research

Key takeaways

1,300+ pull requests merged per week with zero human-written code (up from 1,000+ at first disclosure)
Built on a fork of Block's open-source Goose agent
MCP server ('Toolshed') provides nearly 500 internal tools for context
'Blueprints' orchestration primitive mixes deterministic workflow nodes with agent loops

FAQ

What is Stripe Minions?

Stripe's internal coding agent system that writes code end-to-end and ships over 1,300 merged pull requests per week, with human review but no human-written code.

How do Stripe Minions work?

Engineers invoke agents via Slack, CLI, or web. Agents run in isolated pre-warmed 'devboxes' (standardized AWS EC2 instances), access nearly 500 internal tools via the Toolshed MCP server, and run at most two CI rounds before handing the result to a human, even if failures remain.

What is Stripe Minions built on?

Minions are built on a fork of Block's open-source Goose coding agent, extended with deep internal tool integrations.

Citation review, September 23, 2026: labeled 1,000+ engineer / 10M LOC / under-100 buy thresholds as editorial heuristics, not Stripe figures; qualified the 1,300/week mix via InfoQ; marked the Ramp MCP-depth comparison as unsourced.

Executive Summary

Stripe Minions is the most detailed public case study of enterprise in-house coding agents. As of February 2026, the system produces over 1,300 merged pull requests per week (up from 1,000+ at first disclosure) — with humans reviewing the code but writing none of it.^[1] Stripe says Toolshed, its central MCP server, currently contains nearly 500 tools for internal systems and SaaS platforms.^[1] Per InfoQ, the code Minions touch supports more than $1 trillion in annual payment volume; InfoQ also reports that engineers noted Minions perform best on well-defined tasks such as configuration adjustments, dependency upgrades, and minor refactoring.^[2]

Attribute	Value
Company	Stripe
Type	Internal tool (not for sale)
Foundation	Goose fork (Block)
Public Documentation	February 2026 (Part 1: Feb 9; Part 2: Feb 19)
Headquarters	San Francisco, CA

Product Overview

Stripe Minions are fully unattended coding agents designed for "one-shot" tasks — given a ticket, bug report, or specification, they produce a complete pull request without human intervention during the coding process. The key innovation is deep integration with Stripe's existing developer tooling, allowing agents to work with the same context and tools as human engineers.

Key Capabilities

Capability	Description
End-to-end coding	Writes code from start to finish based on tickets/specs
Multi-surface invocation	Slack (primary), CLI, web UI, internal tool integrations
MCP context	Nearly 500 tools via central "Toolshed" server (docs, tickets, Sourcegraph)
Parallel execution	Multiple minions run simultaneously without git conflicts
Blueprints	Orchestration primitive: deterministic workflow nodes wrapped around agent loops^[1]
Shared rule files	Cursor-format rules synced across Cursor, Claude Code, and Minions

Product Surfaces

Surface	Description	Notes
Slack	Tag agent in any thread to invoke	Primary interface
CLI	Command-line for power users	Secondary
Web	Visibility and management	Administrative
Internal tools	Deep integration with ticketing, feature flags	Stripe-specific

Technical Architecture

Minions run on pre-warmed "devboxes" — standardized AWS EC2 instances that provision in 10 seconds. These environments are identical to what human engineers use, but isolated from production and the internet for security.^[3]

Architecture Flow

Invocation (Slack/CLI/Web)
    ↓
Devbox (isolated EC2 sandbox, 10s spin-up)
    ↓
MCP Server (Toolshed: ~500 tools)
    ↓
Agent Loop (Goose fork)
    ↓
Local lint (under 5s) → CI (max 2 rounds) → Auto-fix
    ↓
Pull Request (human review required)

Key Technical Details

Aspect	Detail
Foundation	Fork of Block's Goose agent, tailored for unattended runs
Execution	Pre-warmed devboxes — standardized AWS EC2 (10-second spin-up)
Context	Toolshed MCP server with nearly 500 tools
Orchestration	"Blueprints" — deterministic workflow nodes combined with agent loops
CI Strategy	Local lint (under 5s), max 2 CI rounds, auto-apply fixes^[3]
Rules	Cursor rule format synced across Cursor, Claude Code, and Minions
Isolation	No production access, no internet

Strengths

Proven scale — 1,300+ PRs merged weekly (as of February 2026), validated in production at one of the world's most demanding codebases; InfoQ reports Minions perform best on configuration, upgrades, and minor refactors, so the weekly count is not evidence that every PR is a large feature change^[2]
Deep integration — MCP server provides agents the same context human engineers have (docs, tickets, build status, Sourcegraph search)
Fast iteration — 10-second devbox spin-up enables rapid feedback loops
Leverages existing infra — Uses same environments as human engineers, reducing agent-specific edge cases
Parallelization — Engineers run multiple Minions simultaneously without git worktree conflicts

Cautions

Requires massive codebase investment — Stripe has hundreds of millions of lines of code and years of devex tooling; this isn't replicable overnight
Human review bottleneck — Agents write code but don't merge; humans must still review everything
Stack-specific patterns — Ruby/Sorbet environment; some patterns may not transfer to other stacks
Dedicated team required — a dedicated internal team maintains the system full-time (team composition not publicly disclosed as of February 2026)
Not for sale — This is internal tooling, not a product you can buy

What Developers Say

Practitioner reaction on the Hacker News thread for Part 2 (131 points) centers on review quality and the lack of concrete examples:

"I am wondering if a lot of the human review is substantive or rubber stamping - as we see with long Pull Requests from humans. I know I would half-ass a review of a PR containing lots of robot code." — 3rodents^[4]

"Code reviews are also an educational moment for seniors teaching juniors as well as an opportunity for people who know a system to point out otherwise undocumented constraints of the system. If people slack on reviews with the agent it means these other externalities suffer." — fnord123

"The glass-half-full here is it's an incredible signal that one of the largest financial gateways in the world is able to do this with current capabilities. Personally, this is exciting." — trevorhinesley

"Where is the detail? Examples? Something concrete? ... Lots of generic statements everyone knows." — gas9S9zw3P9c

The recurring skeptical theme: Stripe has not published example PRs, so outsiders cannot judge whether the 1,300/week figure represents substantive changes or low-stakes mechanical work. No Stripe engineers responded in-thread as of June 2026.

Competitive Positioning

vs. Other In-House Agents

System	Differentiation
StrongDM Factory	Minions require human review; StrongDM eliminates it entirely
Ramp Inspect	Similar architecture; Stripe documents nearly 500 Toolshed MCP tools — a "deeper MCP than Ramp" comparison is unsourced
Coinbase Claudebot	Minions built on Goose fork; Coinbase appears Anthropic-native

When to Reference Stripe Minions

The following scale cutoffs are editorial heuristics, not thresholds Stripe published:

Reference when: Building in-house agents at 1,000+ engineer scale (heuristic)
Consider alternatives when: Team under 100 engineers, no existing devex investment (heuristic)
Buy instead when: Need agents in weeks, not quarters (consider Tembo, Claude Code, Codex) — Tembo is isolated cloud agents with human PR review, a commercial buy-vs-build analog rather than a Minions replica^[5]^[6]

Ideal Customer Profile

This is internal tooling, not a product for sale. However, the architecture is worth studying if you are:

Good fit for similar build (editorial heuristics, not Stripe-published thresholds):

Engineering organization with 1,000+ engineers
Existing devex team (3+ engineers) that can be redirected
Codebase exceeding 10M LOC with proprietary frameworks
Mature internal tooling (ticketing, CI, monitoring)

Poor fit (same editorial heuristics):

Teams under 100 engineers (buy instead)
Standard tech stack without proprietary frameworks
No existing devex investment
Need results in weeks, not quarters

Viability Assessment

Factor	Assessment
Documentation Quality	Excellent (detailed public blog post)
Replicability	Difficult (requires massive devex investment)
Open Source Foundation	Yes (Goose is open source)^[7]
Architecture Maturity	High (production-validated at scale)
Transferability	Medium (principles transfer; implementation is Stripe-specific)

Stripe's detailed documentation makes Minions the reference architecture for enterprise in-house coding agents. While the specific implementation requires Stripe-level investment, the patterns — MCP context, isolated sandboxes, CI integration, Slack invocation — are universally applicable.

Bottom Line

Stripe Minions represent the gold standard for in-house coding agents at elite engineering organizations. The key insight: agents need the same context and tools as human engineers, not a bolted-on integration.

Key metrics: 1,300+ PRs/week (as of February 2026; InfoQ: best on config/upgrades/minor refactors), 10-second devbox spin-up, nearly 500 Toolshed MCP tools (Stripe's figure), max 2 CI rounds. Stripe has not published example PRs.

Architecture pattern: Slack invocation → isolated sandbox → MCP context → CI loop → human review → merge.

Recommended study for: Engineering leaders evaluating build vs. buy for coding agents, devex teams designing agent infrastructure.

Not recommended for: Small teams, organizations without existing devex investment, anyone expecting a product they can purchase.

Outlook: Stripe has defined the reference architecture and is iterating publicly — Part 2 (February 19, 2026) added Blueprints orchestration and cross-tool rule syncing. The buy-vs-build question remains: organizations that cannot staff a Minions-class internal platform can evaluate commercial isolated-cloud agents with human PR review rather than reproducing Stripe's Toolshed and devboxes.

Tembo: Tembo is that buy-side analog — isolated cloud coding-agent sessions, selectable harnesses, and human PR review — not a packaged "Stripe Minions in a box."^[5]^[6] Disclosure: Ry Walker is Tembo's co-founder and CEO.

Research by Ry Walker Research • methodology

Disclosure: Author is CEO of Tembo, which offers agent orchestration as an alternative to building in-house.

Sources