Skip to main content
Unlisted page
This page is unlisted. Search engines will not index it, and only users having a direct link can access it.

Stripe Minions: One-Shot, End-to-End Coding Agents

· 3 min read

Source: Part 1 | Part 2 Author: Alistair Gray (Stripe) Date: Feb 9 & Feb 19, 2026

Key Metric​

Over 1,300 PRs merged per week -- completely minion-produced, human-reviewed, no human-written code.

Why Custom Agents​

  • Codebase: hundreds of millions of LOC across several large repos
  • Primarily Ruby (non-Rails) with Sorbet typing -- rare combo for LLMs
  • Countless homegrown libraries specific to Stripe
  • Processes >$1T/year in production
  • Philosophy: "if it's good for humans, it's good for LLMs, too"

Entry Points​

  1. Slack (primary) -- tag the Slack app from any thread, full conversation context included
  2. CLI and web interfaces
  3. Internal tool integrations -- docs platform, feature flags, ticketing
  4. Automated triggers -- CI detects flaky tests -> auto-ticket with "launch minion" button

Monitoring & Output​

  • Web interface for real-time observation of agent decisions
  • On completion: branch, CI push, PR following Stripe's template
  • If code is good -> open PR for colleague review
  • If not -> provide additional instructions, minion pushes updates
  • Partially correct output used as foundation for focused human work

Devboxes (Cloud Dev Infrastructure)​

PropertyDescription
ParallelizabilityMultiple agents on separate tasks simultaneously
PredictabilityStandardized configs ensure consistent behavior
IsolationWork confined to individual environments
  • "Cattle, not pets" -- standardized, easy to replace
  • Warm pool achieves "hot and ready" in ~10 seconds
  • Pre-cloned repos, cached services
  • Isolated from production and internet

Agent Harness: Forked Goose​

Internally forked Block's Goose (open-source coding agent) in late 2024. Key distinction: minions operate without human supervision -- full permissions, no confirmation prompts, safe within isolated devboxes.

Blueprints: The Orchestration Framework​

Central architectural innovation. A "state machine that intermixes deterministic code nodes and free-flowing agent nodes."

Example node types:

  • Agentic nodes: "Implement task," "Fix CI failures" (free-form LLM reasoning)
  • Deterministic nodes: "Run linters," "Push changes," git ops, testing (guaranteed execution)

This hybrid approach reduces token consumption and improves reliability.

Context: Rule Files​

  • Directory-specific and pattern-based rules (not global -- would overwhelm context)
  • Standardized on Cursor's rule format
  • Synchronized across minions, Cursor, and Claude Code

Context: MCP (Model Context Protocol)​

"Toolshed" -- centralized internal MCP server with ~500 tools for internal systems and SaaS.

  • Minions receive an intentionally small subset by default
  • Per-user customizable additional tool sets
  • MCP is the common language across all Stripe agents
  • Deterministic pre-execution of relevant MCP tools for context hydration

Feedback Loop​

  1. Local linting -- heuristic-based, <5 seconds per push
  2. CI selective testing -- from 3M+ tests, only relevant ones run
  3. Autofixes -- many tests include autofixes, applied automatically
  4. Single retry -- one resolution attempt for remaining failures

Hard rule: "often one, at most two, CI runs" -- shift feedback left.

Security​

  • Devboxes in QA environments
  • No access to production services or real user data
  • Internal control frameworks preventing destructive actions

Stripe Minions Architecture