We are stuck in the "Chatbot Era." You type a prompt, you wait, you get text back. It's synchronous, human-in-the-loop, and fundamentally limited.
The next phase of AI is Agentic. It's asynchronous, autonomous, and tools-based.
The Four Patterns of Agency
Andrew Ng recently popularized four design patterns that define agentic workflows. If you are building AI apps today, memorize these:
1. Reflection
The model critiques its own output. "Review the code I just wrote for security vulnerabilities." It's a loop, not a line.
2. Tool Use
The model recognizes it doesn't know the answer and calls an external function (Search, Calculator, API). This solves the hallucination problem for deterministic tasks.
3. Planning
The model breaks a high-level goal ("Build a website") into sequential steps ("Write HTML," "Write CSS," "Deploy").
4. Multi-Agent Collaboration
This is where it gets crazy. One agent acts as the "Coder," another as the "Reviewer," and another as the "Product Manager." They debate each other until the output is refined.
"Future software won't be written; it will be negotiated between agents."
The Architecture Shift
This changes how we architect backends. We aren't building REST APIs for humans anymore; we are building robust tool definitions for LLMs. The best API documentation is now the most important feature of your product.