Skip to content

Blog

1,300 agent PRs a week rest on three million tests

Stripe's 1,300 unattended pull requests a week get quoted everywhere. What makes them work is the test layer and the harness underneath the agents.

Written by

Specification and review

Published: 9 June 2026 Updated: 2 min read

Stripe ships 1,300 unattended PRs a week. The number gets quoted everywhere. I think the wrong part of the number is what gets quoted.

What makes the 1,300 work isn’t the agents. It’s the 3 million tests sitting underneath them.

Teams copy the surface

That’s the part most teams skip when they read the case study and try to copy it. They look at the surface, agents writing PRs, and try to install agents on top of a codebase that has, generously, 30k tests and a flaky CI. The result is the loud kind of failure.

The stack reads like industrial infrastructure

Their stack is built like a plant:

  • A large Ruby codebase.
  • Three million tests.
  • A blueprint engine that interleaves deterministic code with non-deterministic agent loops, so the agent runs inside a frame the system already trusts.
  • A pool of pre-warmed devboxes that spin up in 10 seconds, so an agent never sits idle waiting for an environment.
  • A tool shed of around 500 MCP tools, with each agent handed only the subset its task needs.

That isn’t a coding agent. It’s a coding factory. The PR is the output. The 1,300 is what falls out at the end of the line.

Reinforce the factory before you add agents

I’d say the move that’s transferable isn’t “use agents.” It’s “pick the part of your factory that needs the most reinforcement before you scale up the agent count.” For most teams I work with, that’s the test layer. You can ship two agents on top of 3M tests. You cannot ship 200 agents on top of 30k.

The harness matters more than the model

This is also why the harness conversation matters more than the model conversation right now. The model is a parameter. The harness is the system the model runs in.

Most teams are still picking models. That company spent the last few years picking the harness.

What’s the part of your harness you’re underinvesting in right now?

(Source: Stripe engineering, “Minions: Stripe’s one-shot, end-to-end coding agents”, parts 1 and 2, February 2026.)

Further reading

  1. Refinement, not coding, was the bottleneck all along Several teams turned the same rough story into wildly different plans. The agent did not create the ambiguity. It showed that deciding what the work means was always the slow step. 2 min readShort post
  2. When code is cheap, agreeing on what to build is the bottleneck Slow typing used to force teams to agree before they built. With agents, a vague ticket produces two confident builds, so the fix moves upstream into a short alignment doc. 2 min readShort post
  3. When code is cheap, the design decision is the expensive part When a full feature costs a few dollars to generate, throwing code away stops being waste and becomes the cheapest way to test an idea. The constraint moves from writing the code to deciding what survives. 1 min readShort post

All 23 posts on Specification and review

ClientsClient work

Test the constraint in your system

The Factory Readiness Assessment connects this research to your value stream, your data, and the next decision.

Training on this topic: Agentic Engineering for Developers and QA
See the assessment

Want to check one team first? The self-check: twelve yes or no questions, three minutes, instant result, no account. Take the self-check