Not everything in AI right now is as new as it sounds. Every week there’s a fresh term making the rounds ... “agentic workflows,” “context engines,” “MCP servers.” But when you peel it back, most of these are just evolutions of the same core ideas the industry has already been refining over the last few years: prompt engineering, RAG, and tool use (whether you called them plugins, functions, or skills a year ago).

What’s actually changed is that models are getting better, the patterns are maturing, and the ecosystems are starting to settle into more consistent shapes. The tech itself is solid, but most of what’s getting hyped is just improved packaging and orchestration.

Even the idea of an “AI agent” usually boils down to a loop: retrieve some info, call a tool, maybe score a few options, then keep going (to be fair, there is a bit more nuance to this one that's worth talking about in a separate conversation 😅). Same goes for frameworks like langchain and semantic kernel; incredibly useful things but not magic if you understand the elemental building blocks they're gluing together.

And these aren’t just engineering conversations anymore. I’m hearing the same questions from non-technical folks as well ... everyone's trying to navigate what’s real, what’s useful, and what’s just branding.

A few other terms I’ve seen lately that feel like this:

• "Memory": state tracking plus a vector store
• "Cognitive architecture": prompt templates with logic and tool use
• "Copilots": chat + retrieval + API calls
• "Synthetic users": prompted chat with constraints
• "Reasoning engine": Chain of Thought prompts with some heuristics thrown in
• "MCP": basically tool calling plus JSON-RPC with some standardization

This isn’t a knock on any of it. There’s a lot of powerful stuff happening. But it helps to stay grounded and recognize when something is just a new wrapper on increasingly familiar tech.