Type: url
Original: https://www.humanlayer.dev/blog/12-factor-agents
Status: partial
Summary
A HumanLayer blog post, published April 3, 2025 and based on the original 12 Factor Agents write-up on GitHub, that proposes twelve principles for building LLM-powered software reliable enough for production customers, in the spirit of the 12 Factor App methodology. It argues that good agents are mostly deterministic software with LLM steps placed at the right points, and that long ‘loop until done’ agents get lost as their context grows. The factors cover turning natural language into tool calls, owning prompts, context windows and control flow, unifying execution and business state, pause/resume APIs, contacting humans via tool calls, compacting errors, small focused agents, triggering from anywhere, and treating the agent as a stateless reducer.