Stripe Treats Agents as a Factory: 20 Days to 4 Days; What's Valuable Is the 125 Old Projects

Stripe publicly shared its internal Agent factory. For the same payment integration task, ordinary Claude Code took 20 days, while their process took 4. Payment integrations used to take up to 6 months; now they are compressed to 2 to 6 weeks.

In the tweet, Yarchi broke down five actions, and the core is very concrete.

1. One prompt per step. They studied more than 125 historical integrations and turned each step into a reusable prompt. There are now more than 100 of them.
2. A second agent watches the first agent. One agent does the work, and another reads its transcript, recording where it got lost, what it relearned, and which tools failed. Engineers use these records to revise prompts.
3. Prompts update themselves. After each run, learning logs, decision logs, and review comments are left behind. Another agent turns them into prompt fixes, some of which can be applied with one click.
4. Run agents with code. They first tried an AI orchestrator, but it was slow and expensive, and sometimes it did the work itself instead of delegating it. Later they rewrote it as ordinary code. A commenter summarized it more directly: LLM makes judgments, boring code does orchestration.
5. Agents test themselves. First run test payments in a sandbox, then have a human review the PR.

There are only three steps to copy the method: save every effective prompt, have a second agent review the transcript, and fix prompts rather than outputs.

But don't rush to copy it. In a reply, commenter John Rood pointed out a prerequisite: the gains come from decomposition, and the 125 historical integrations made the step list discoverable instead of having to invent it from scratch. Copy the harness without copying the history, and the result may be the 20-day version plus more moving parts. So-called compounding means the 100th run reads what the 1st run learned. Commenter Road N also said that the same model was used both times, and the difference lies in the written-down steps. Another commenter, Jatin Garg, warned that changing a shared prompt could break integrations that already work, so old payment methods must be tested before rolling out fixes. AI Ferryman Alex asked how to prevent library bloat and expiration after more than 100 saved prompts.

In another article by Yarchi about Agent Harness, this outer layer is broken into six things: how the loop stops, the tool menu, memory, what is left behind after a crash, permission boundaries, and who decides completion. The cheapest starting version has five items:

1. An instruction file under 100 lines, used as a documentation directory, not written as an encyclopedia.
2. If the same mistake appears twice, write it as a linter. Don't write prose asking the model to pay attention.
3. Put four files in the repository: SPEC.md, PLAN.md, PROGRESS.md, DECISIONS.md. Commit after every working change.
4. Pin down rules that must never be summarized away. Security, policy, hard limits.
5. Prepare 20 tasks, run each three times, and score based on the worst run.

The article also cites a cost comparison: Anthropic's harnessed run took 6 hours and $200, while the unharnessed version took 20 minutes and $9. More than twenty times the price bought better results. This account needs to be calculated clearly first. Harness is suitable for work that originally could not be delivered at all, not for tasks where you just want to save 20 minutes.

Agent harness cannot be avoided. You already have one right now, piled up from 2 a.m. workarounds, rules pasted into config and forgotten, and permission prompts approved because clicking was faster than reading. All these decisions happened, just without design. Write the stopping rules first, then put in those four files. It can be done in two afternoons and can survive the next model release.

Sources: Yarchi's tweet, Stripe's blog, How to Design an Agent Harness.