Over the past three years, the way we interact with large language models has undergone a profound transformation. Initially, users simply asked questions and received answers. Today, the paradigm has shifted to 'harnesses'—orchestrated systems where multiple agents, tools, and workflows collaborate to achieve complex tasks. This evolution reflects a maturation of the AI ecosystem, moving from novelty to utility. For developers and technical founders, understanding this shift is crucial: building with LLMs now requires designing for autonomy, error handling, and multi-step reasoning rather than just crafting the perfect prompt. The post explores key milestones, including the rise of agent frameworks, function calling, and retrieval-augmented generation, and offers insights into what the next phase might look like. It serves as a valuable signal for anyone building AI-native products or integrating LLMs into production systems.
A retrospective on how LLM interaction patterns have evolved from simple prompts to agentic harnesses, with implications for developers.