# gray-stripe-minions-coding-agents-part2-2026-02-19

## Veille

Stripe's Minions — end-to-end coding agents, devboxes, blueprints, and hybrid orchestration at scale

## Titre Article

Minions: Stripe's one-shot, end-to-end coding agents -- Part 2

## Date

2026-02-19

## URL

https://stripe.dev/blog/minions-stripes-one-shot-end-to-end-coding-agents-part-2

## Keywords

coding agents, Stripe, Minions, devboxes, blueprints, hybrid orchestration, state machine, MCP, Toolshed, agent security, CI/CD, developer productivity, goose, agent infrastructure

## Authors

Alistair Gray

## Ton

**Profile**: Perspective of an internal practitioner engineer, technical and descriptive register, high technical level. Alistair Gray writes from within Stripe with the legitimacy of someone who builds and operates the system described. **Style**: Methodical engineering prose, structured into successive architectural components (devboxes, harness, blueprints, context, feedback). Concrete operational metaphors — "cattle not pets" for devboxes, "putting LLMs in contained boxes" for blueprints. Factual and pragmatic tone, little speculation, lots of numbers (1,300+ PRs/week, 10 seconds, 3M+ tests). Authority comes from measured results and deployment scale. Target audience: platform engineers and enterprise AI infrastructure leads.

## Pense-betes

- **Dramatic scaling**: from 1,000 to 1,300+ PRs merged per week between Part 1 and Part 2
- **Devboxes = "cattle not pets"**: disposable AWS EC2 instances, pre-warmed pool, ready in 10 seconds, same environment as human engineers — complete isolation
- **Agent harness**: custom fork of goose (Block), optimized for unassisted execution — no confirmation prompts thanks to devbox isolation
- **Blueprints = hybrid orchestration**: state machine that mixes deterministic nodes (traditional code) and agent nodes (free-form LLM loops). Saves tokens and improves reliability
- **Key philosophy**: "putting LLMs in contained boxes" rather than giving them total free rein
- **Unified context**: rules files standardized on the Cursor format, synchronized to the Claude Code format. Same rules for Minions, Cursor, and Claude Code
- **Centralized MCP via Toolshed**: about 500 tools, curated subsets per agent — avoids context overload
- **Security**: devboxes in QA-only environment, no production access, no user data
- **Two-stage feedback**: pre-push lint hooks (<1 sec), then 1-2 CI cycles max against a battery of 3M+ tests
- **Strategic insight**: investments in human developer productivity pay dividends for agents too — same tooling, same infrastructure

## RésuméDe400mots

Alistair Gray details in this second part the technical implementation of Minions, Stripe's end-to-end coding agents, which now merge more than 1,300 pull requests per week, up from the 1,000 announced in Part 1.

The infrastructure relies on devboxes — ephemeral AWS EC2 instances treated as "cattle, not pets." A pre-warmed pool allows each agent to get a complete environment in 10 seconds, identical to that of human engineers. This total isolation is fundamental: it allows agents to run without supervision or confirmation, since they cannot access production or user data.

The agent harness is a custom fork of goose, Block's open source tool, optimized for unassisted operation. Confirmation prompts have been removed because devbox isolation makes this safeguard redundant.

The major architectural innovation lies in blueprints — a hybrid orchestration system combining deterministic nodes and agent nodes within a state machine. Predictable steps (cloning a repository, creating a branch) are executed by traditional code, while creative steps (understanding a problem, writing code) are delegated to LLM loops. This approach "puts LLMs in contained boxes," saving tokens and significantly improving reliability compared to a purely agentic approach.

For context, Stripe has standardized its rules files on the Cursor format, with automatic synchronization to the Claude Code format. The same rules serve Minions, developers using Cursor, and those using Claude Code — a unification that amplifies the return on investment of every context improvement.

Tooling is centralized via Toolshed, an MCP server exposing about 500 tools. Each agent receives a curated subset suited to its task, avoiding the context overload that degrades LLM performance.

The feedback loop operates in two stages. First, pre-push lint hooks run in under a second to catch trivial errors. Then, the agent submits its code to one or two CI cycles maximum against a battery of more than 3 million tests.

Gray concludes with a cross-cutting philosophy: the investments Stripe has made to improve the productivity of its human engineers — tooling, infrastructure, CI — now pay double dividends by also serving agents. This is a powerful argument for organizations hesitant to invest in their development platform.

## GrapheDeConnaissance

- Stripe —a_créé→ Minions (TECHNOLOGIE, 0.98)
- Minions —mesure→ 1300+ PRs fusionnées par semaine (MESURE, 0.95)
- Minions —utilise→ devboxes (TECHNOLOGIE, 0.97)
- Devboxes —est_basé_sur→ AWS EC2 (TECHNOLOGIE, 0.95)
- Devboxes —mesure→ disponibilité en 10 secondes (MESURE, 0.93)
- Minions —utilise→ goose (TECHNOLOGIE, 0.96)
- Block —a_créé→ goose (TECHNOLOGIE, 0.9)
- Blueprints —utilise→ nœuds déterministes et nœuds agents (CONCEPT, 0.97)
- Stripe —utilise→ format Cursor (règles) (TECHNOLOGIE, 0.92)
- Stripe —a_créé→ Toolshed (TECHNOLOGIE, 0.94)
- ~500 outils MCP —fait_partie_de→ Toolshed (TECHNOLOGIE, 0.93)
- Devboxes —observé_dans→ environnement QA (CONCEPT, 0.95)
- Alistair Gray —affirme_que→ investissements productivité humaine profitent aux agents (AFFIRMATION, 0.9)
- Blueprints —améliore→ fiabilité agents (CONCEPT, 0.92)

---
Canonical: https://www.thekb.eu/en/fiches/gray-stripe-minions-coding-agents-part2-2026-02-19/