Stripe Minions: unattended blueprints beat shoulder-surfing Claude

Stripe’s Feb 19 Part 2 on Minions is a receipt for always-on coding under real merge load: 1,300+ PRs/week are completely minion-produced (human-reviewed, zero human-written code) — up from ~1,000 in Part 1. The developer-facing pitch is FOMO-shaped: attention is scarce, so unattended agents parallelize the one thing a human cannot — watching the session. The transferable engineering is the harness, not the Slack button.

Minions run on hot-ready isolated “devboxes” (pre-warmed EC2 clones of Stripe’s human environment, ~10s spin-up) so agents get parallel, predictable working trees without laptop shadow context, personal SSH inheritance, or worktree juggling. The orchestration primitive is a blueprint: a coded state machine that interleaves free agent nodes (“Implement task,” “Fix CI failures”) with fully deterministic nodes (“Run configured linters,” “Push changes”) that never call an LLM. Local lint/heuristics shift feedback left before CI; the blueprint then allows at most two CI rounds (autofixes applied; one agent repair pass) and returns the branch to a human — diminishing returns beat infinite repair loops. Context is subdirectory-scoped Cursor-format rules plus a curated slice of Toolshed (hundreds of MCP tools, small box per agent). Quarantined QA-env blast radius is what lets them skip confirmation prompts.

Why it matters for code review

Sustained overnight and background throughput only compounds when verification is wired into the loop as code, not as a morning vigil. Stripe’s blueprint pattern is the industrial form of modest verified momentum: keep agents running, force deterministic gates and bounded CI tax before the PR queue, and leave human review for merge judgment — not for babysitting a chat. Peak-session Claude mastery without an isolated environment and a halt policy just dumps larger diffs into the same attention bottleneck Minions were built to relieve.