Seven agentic AI patterns that hold up in production, the failure each one fixes, and when a plain script beats an agent entirely.
Seven patterns cover nearly every agentic AI system I have seen survive contact with production: the single agent in a tool loop, prompt chaining, the router, orchestrator-workers, the evaluator-optimizer loop, human-in-the-loop approval gates, and the memory-augmented long-running agent. Everything that ships is one of these or a combination of them. Most of what does not ship is a ten-agent diagram that looked spectacular on a whiteboard. This is the numbered, citable version of a conversation I have with clients every month. The vocabulary is now stable across the industry: Anthropic's Building effective agents named most of these patterns, OpenAI's practical guide to building agents converged on the same shapes, and LangGraph turned them into framework primitives (durable execution, routing, human interrupts). What is still missing in most writeups is the failure mode of each pattern, and that is where projects die. I build these systems for SMEs, and I run…