humanlayer/12-factor-agents

Provides a methodology for building production-grade LLM agents through 12 design principles—covering prompt ownership, context window management, human-in-the-loop integration via tool calls, and stateless reducer patterns. Emphasizes deterministic software with strategic LLM integration rather than generic agentic loops, prioritizing explicit control flow, execution state unification, and pause/resume capabilities. Inspired by 12factor.net, the guide targets engineers building customer-facing AI systems and complements frameworks like LangGraph through architectural best practices rather than implementation specifics.

18,686 stars. No commits in the last 6 months.