What is a "cognitive architecture"?

The article defines a "cognitive architecture" as how an LLM-powered system thinks: the flow of code, prompts and LLM calls that takes user input and produces actions or a response. It describes five levels of autonomy, from fully hard-coded code, to a single LLM call, to a chain of LLM calls, to a router where the LLM chooses actions, to a state machine combining routing with a loop, and finally to autonomous agents, where the system itself decides which steps and instructions are available. The author says none of these architectures is strictly better than the others, and that developers should experiment with them as often as with prompts. The article credits Flo Crivello with coining the term and promotes LangChain and LangGraph as frameworks for this work.