This year, we’re seeing a noticeable shift at the AI application layer—from workflow-based automation toward more autonomous agents. Tools like n8n and Dify represent the former, while OpenClaw, Hermes, and CoWork are early examples of the latter. That said, new paradigms rarely replace old ones outright. More often, they evolve by absorbing and reshaping what came before—a kind of “sublation” rather than simple replacement.
In my previous article on our new technology stack AmphiLoop, I introduced a new programmatic architecture called “Amphiflow”. Amphiflow is what I believe to be the first truly “decision–execution decoupled” agent architecture. It combines workflow and agent paradigms into a single runtime, with the ability to switch between them automatically.
I’m very excited to launch a brand-new open-source repository—AmphiLoop. But this is more than just a code repository; it’s a brand-new methodology, tech stack, and toolchain for building automation systems & AI agents. It allows us to describe and orchestrate tasks using natural language, while also having the ability to automatically switch between workflow mode and agent mode at runtime.
This is precisely where the name “AmphiLoop” comes from—Amphibious Loop! Next I’ll go into detail about the whole system and the principles behind it. The article is long, and was entirely written by my own hand (with no AI assistance). I recommend bookmarking it first and reading it slowly later.