Intelligent Agents in AI: Building Blocks, Workflows, and Agents
So, we've been talking about how focusing on the capabilities of AI agents is more important than getting bogged down in all the different tools available to build them. It's easy to get overwhelmed, right? There are so many frameworks popping up daily, like CrewAI, Pydantic AI, Autogen (AG2), Dify, and n8n, it is hard to tell which is best. Turns out, many successful real-world applications of AI agents don't even rely on these complex frameworks. They're built using simple, composable patterns.
But it begs the question: what are these "intelligent agents" everyone's talking about? What is an "intelligent agent in AI," specifically?
Think of it this way: intelligent agents are like super-smart virtual assistants. But, unlike the ones on your phone, these can be designed to do way more than just set reminders or tell you the weather.
The term "agent" can be defined broadly. Some define agents as fully autonomous systems that independently accomplish complex tasks over extended periods, utilizing various tools. Others refer to more prescriptive implementations following predefined workflows.
There are 2 types of agentic systems to differentiate: workflows and agents.
Workflows are like detailed instruction manuals. They tell the AI exactly what steps to take, in what order, using what tools. The Large Language Model (LLM) at the heart of the system follows a pre-defined code path. It's very predictable.
Agents, on the other hand, are more like seasoned experts. They have a goal, and they figure out the best way to reach it, choosing the right tools and steps along the way. The LLM controls its processes and tool usage, maintaining control over task completion.
When to Use Which: Workflows vs. Agents
Choosing between workflows and agents is like choosing between a detailed recipe and a chef's intuition.
Workflows are great for tasks you do repeatedly that always have the same steps. Think of processing invoices or generating reports. You want consistency and predictability. Many times, optimizing individual LLM calls with retrieval and in-context examples is sufficient.
Agents are better for situations where you need flexibility and problem-solving. Imagine you need to research a new market or come up with creative marketing ideas. The agent can adapt to new information and make decisions on the fly.
The Bottom Line: Start simple! You might not even need a full-blown agent. A well-crafted prompt for an LLM might be all it takes. If you do need more, workflows are often the next logical step before diving into complex agent architectures. They're simpler to build and understand.
Frameworks: Helpful, But Not Always Necessary
Now, about those frameworks like LangGraph, Amazon Bedrock's AI Agent framework, Rivet, and Vellum – they can be helpful. They simplify things like calling the LLM and defining tools. But remember the successful applications that don't even use these frameworks? Plus, they can add a layer of complexity that makes it harder to see what's happening "under the hood."
My advice: If you're just starting, try building directly with the LLM API first. You'll learn a lot, and it's often easier than you think. Many patterns can be deployed in just a few lines of code. If you do use a framework, take the time to understand how it works.
And here's a pro tip: If you're looking to build any no-code application that integrates powerful LLMs like Gemini, ChatGPT, Anthropic,... check out Diaflow 🤖. It's designed to help you create advanced, cutting-edge LLM applications with ease, no matter your level of experience. Think of it as your secret weapon for building the next generation of intelligent tools.
In the next post, we'll break down the building blocks of workflows and agents even further. We'll look at the core components of any intelligent agent in AI, and start to see how you can design your own! Stay tuned!"