Stripe Shows Its Cards: Minions Blueprint Architecture Uses 400 Tools, Deploys ~15 Per Run, and Ships 1,300+ PRs Weekly

Stripe published Part 2 of its Minions engineering blog, and it's the most detailed production agent architecture I've read from any company this year. The numbers alone are worth the read: 1,300+ weekly merged PRs from coding agents. But the architecture decisions matter more than the throughput.

Here's what caught me off guard. Before any LLM runs, a deterministic orchestrator prefetches all the context the agent will need. It scans Slack threads for links, pulls Jira tickets, and runs Sourcegraph MCP queries. Then it curates roughly 15 tools from a central "Toolshed" MCP server that contains 400+ available tools. The LLM never sees all 400. It gets the 15 that are relevant.

This is the opposite of the "give the agent everything and let it figure it out" approach I see in most open-source agent frameworks. Stripe's insight: more tools degrades performance. Targeted tool curation saves tokens and produces better results. LangChain's own data backs this up, showing accuracy jumps from 17% to 92% with progressive skill disclosure.

The Blueprints pattern alternates between deterministic nodes (guaranteed execution, no LLM calls, no token spend) and agentic loops (LLM reasoning for ambiguous decisions). Each minion runs in an isolated, pre-warmed devbox that spins up in 10 seconds, with zero internet access and zero production access. That isolation lets them run infinite parallel agents safely.

I've been building something similar on a smaller scale, using deterministic context prefetching before Claude Code sessions. The pattern works. You spend a few hundred milliseconds gathering context upfront and save minutes of agent wandering. The lesson from Stripe: the hybrid deterministic-plus-agentic pattern is how you build production agents that actually ship code at scale. Pure autonomy is a demo. Structured orchestration is production.

For builders: steal this pattern. Build a small tool catalog, curate per-task, and front-load your context gathering with deterministic code. Don't let your agent discover what it needs. Tell it.