The AI industry is undergoing a massive shift. We are moving from models that just “talk” to models that actually “do”. This shift is powered by Agentic Workflows.
What Makes a System “Agentic”?
Dr. Andrew Ng recently popularized four key design patterns that define agentic workflows. By implementing these patterns, even older models like GPT-3.5 can routinely outperform GPT-4 operating in a standard zero-shot chatbot mode.
1. Reflection (Self-Correction)
Instead of accepting the first thing the LLM writes, an agentic system asks the LLM to critique its own work. “Here is the code you wrote. Here are the common bugs associated with this logic. Review the code and rewrite it fixing any issues.” This simple feedback loop drastically improves output quality.
2. Tool Use (Function Calling)
LLMs are terrible at math and don’t have real-time data. Agentic systems solve this by giving the LLM tools. If the user asks for the weather, the LLM pauses, writes a JSON payload to call a Weather API, and then uses the response to answer the user. Tool calling is the foundation of AI Engineering.
3. Planning
For complex requests, the agent breaks the goal into smaller steps before acting. It creates a task list: 1. Search web for X. 2. Summarize findings. 3. Draft email. It then executes these steps sequentially.
4. Multi-Agent Collaboration
Different LLMs are given different personas. For example, one agent acts as a “Coder” and writes software. Another agent acts as the “QA Tester” and tries to break the code. They converse back and forth autonomously until the code passes all tests.
How Do We Build Them?
Building agentic workflows requires specialized frameworks that manage “State” (the memory of what has happened so far). The industry standard for this is LangGraph.
LangGraph allows you to define nodes (agents or tools) and edges (the logic that dictates where the flow goes next). This allows you to build highly reliable AI applications with safe failure modes and “human-in-the-loop” approval steps.