Stripe's minions merge 1,300+ pull requests a week with no human-written code. Sierra's Pinecone opens 70% of their PRs. Last week those numbers anchored two good essays arguing opposite conclusions: one says you probably don't need to build your own, one says you probably should.
For most teams the answer you should use Paperclip - which already has thousands of commits from outside contributors - instead of spending two quarters building your own (which you then have to maintain). Here's my case:
- It's four large systems
Read the Stripe and Sierra write-ups and count the distinct systems they describe:
A runtime that executes agents in sandboxes.
A work system that takes in tasks, routes them, and gets the output reviewed.
A permission system that decides what each agent may touch and who answers for it.
A learning system, where every failure becomes a rule, a skill, or a check that the next run inherits.
The build-versus-buy essays are a debate about the first system. The other three barely come up. But it's all of them together that turns “an agent in a VM” into 1,300+ merged pull requests a week.
Paperclip already provides all four:
The place where human intent is declared and the work is verified - an agentic task manager where approval is part of the task, not a Slack message next to it.
The org chart that defines permissions for every agent and keeps track of which human is responsible for each one.
Built-in agent training, plus a company-wide audit log that keeps every trace in one place - which is what makes company-wide learning possible at all.
Infrastructure that runs any model from any provider, battle-tested by hundreds of open-source contributors rather than just your engineering team.
An agentic task manager, an org chart that drives your agents, agent employee training, and the OS underneath - in one product. The task manager, permissions, and tools are built into the same system as your org chart. Your evals run in the same system your agents work in. You don't spend your week wiring four tools together. Code is free now; verification is not.