Why I’m Learning About Agentic AI
The world has got used to Generative Pre-trained Transformer (GPT). It has transformed how human-machine interaction works. Now we can use natural language to command machine, AI in this context, so that the interaction mimics human-to-human interaction. Then, it was called Graphical User Interface (GUI), which we learned about in the Human-Computer Interaction topic, to make software intuitive by design.
So, why I’m learning about agentic AI? As Large Language Models (LLMs) grow, I questioned myself how could it be useful in everyday’s works. I used to automate my work so I could offload my routine tasks and simply monitor them. Automation really comes in handy, even before it was reintroduced as Robotic Process Automation (RPA). We may now call it an agentic workflow. This so-called agentic workflow is an automation process for tasks that can be delegated to an AI agent.
Before Agentic AI
First, it was simply traditional software. I was a software engineer by training. Then came the Machine Learning era over the past decade, which we now call traditional or classical AI. I shifted my focus to ‘Data Management’ at the right time. I’ve done end-to-end work on a data pipeline as a Data Product Manager: Data Engineering to Data Science to Data Architecture.
Next, the “current AI” boomed! I had to hone my technical skills again because it’s different from just a machine learning model. I learned about transformers, LLMs, and AI agents for the past year. It does evolve quickly.
So, What is an AI Agent? What is an Agentic Workflow?
An AI agent uses an LLM as its brain to reason about a goal, decide what to do, use tools, observe the results, and act accordingly. An agentic workflow is a set of tasks that we delegate to agents. Yes, the results will no longer be deterministic in the agentic workflow context, unlike the old automated workflow. The AI agent can plan and act using the provided tools, such as web search, functions, or other LLMs, based on their reasoning.
Why am I Excited about Agentic AI?
It has potential. For example, in data work, the agents can build the query on their own, help with the analysis and visualization, and even give an explanation in natural language that is easy to interpret. The question remains how good is the results from the agent? If an agent takes an action, how can we know that it reliably takes the right action (especially for a customer-facing agent)?
From One Single Agent to Many
If one agent can be used for multiple tasks, why do we need multiple agents? It is the same as human beings; we need multiple experts so each agent needs to be trained to have its own specialization to give more reliable results. But, does adding more agents eventually make a system better? How to evaluate that? Besides, more agents also mean more cost, more latency, more prone to error, and more communication and coordination problems among the agents. We have to rebuild the entire system for agent-to-agent interaction.
What I Want to Explore?
Many questions in mind right now. I might not yet the experts. Hence, it is my platform to learn and share my learning process. Some questions I wan to investigate:
- When does an agent outperform a conventional workflow?
- When does a multi-agent system outperform a single agent?
- What kinds of tasks benefit from specialization?
- How should agents learn from real-world behaviour?
- How do we evaluate whether simulated agent behaviour resembles reality?
- Can real behavioral data make agents more reliable?
- Can these systems be tested before they interact with real customers?
My next step is to move from understanding agents conceptually to building one. I’ve built multi-agent systems for an AI Studio course last spring, but I would want to understand them deeply and explain what I am doing here in this journal. I’ll be working on a project called Agent Reality Lab as part of my learning and experimentation with the intersection of multi-agent systems, simulated environments for agents, and real human behaviour.