A technical deep dive into Stripe's reported 'Minions' architecture, which uses one-shot (single-turn) LLM agents instead of multi-step agentic loops. The core thesis is that investing in context engineering before a single LLM call yields better reliability, lower latency, and lower cost than chaining multiple reasoning steps
Table of contents
Table of ContentsWhy Bet Against Agentic LoopsWhat Are One-Shot 'Minions'? Core Concepts ExplainedArchitecture Deep Dive: How Minions Work at ScaleImplementing a One-Shot Agent System in Node.jsBuilding a Simple Dashboard in ReactWhen to Use One-Shot Agents vs. Multi-Turn AgentsImplementation Checklist and Production ConsiderationsSimplicity as a Scaling StrategySort: