TL;DR: Stripe detailed an internal agent setup that cut a repeatable integration task from 20 days to 4, and the reusable parts are prompt-per-step plus a reviewer agent.
The Catalyst
Per the shared write-up, Stripe turned 125+ past payment integrations into 100+ reusable per-step prompts, with a second agent reading the builder agent's transcript to log where it got lost and which tools failed. They replaced an LLM-based orchestrator with plain code after it proved slow, expensive and prone to doing the work itself instead of delegating; integration time went from up to 6 months to 2-6 weeks, and agents run sandbox test payments before human PR review.
The Blast Radius
Anyone running a single long agent prompt on repeated work loses the two things Stripe credits: per-step prompts you can edit independently, and a transcript reviewer that produces prompt fixes instead of output patches. Check your own loop for the deterministic-code-vs-LLM-router decision first, then whether failed runs leave logs that change a prompt rather than a result.
Source: x_com