Are AI Agents the Future of Gen AI? - Dream

Hriday Jain
Mar 14
6 min read

Introduction

As promised in my last blog, I’m back with probably one of the hottest topics in the AI world — Agentic AI! Just a few years ago, LLMs like ChatGPT were first introduced, they blew everyone’s minds, right? People were all like, "LLMs can do this, do that!" But now? They’re like, "Meh." And the new talk of the town? It’s all about AI agents, and it should rightfully be so! Over the past month, I’ve been deep into researching AI agents, and the more I learned, the more I realized—one blog just wouldn’t cut it. So, instead of cramming everything into a single post, I’m turning this into a three-part blog series: Dream, Design, and Deploy ( I like to call them the 3 D’s of Agentic AI ). Each post will break down a different stage of AI agents, and it all starts now with Dream!

What are AI Agents?

Whenever I talk to people about the AI revolution, I get the same predictable response: "Oh, GenAI is great, but it can never replace me. Maybe the way we work will change a bit, but that’s okay." And I don’t agree with this one bit. GenAI can certainly make life easier, but the belief that it can never replace humans only holds true for simple chat-based LLMs and SLMs—not AI agents. Agents aren’t just for back-and-forth conversations—they autonomously complete tasks.

I always tell people: if you really want to understand the magnitude of Agentic AI and the impact it can create, you first need to accept a fundamental reality—AI agents are a vastly superior workforce, capable of completing tasks more efficiently and accurately than any human employee.

This is the fundamental difference between chat-based LLMs and AI agents. Chatbots like ChatGPT require continuous prompting, and even then, they can only assist in completing a task—they can’t complete it independently. Think of it like this:

An LLM is like a consultant—it gives you advice and insights, but you still have to do the work.
An AI agent is like a full-fledged employee—it analyzes, decides, and executes, often better and faster than a human.

Now, don’t go thinking that LLMs and AI agents are two completely different things—they’re not. An agent is basically just an LLM with extra powers. In simple terms, it’s an LLM trained to respond in a specific way and given external API accesses to interact with the world and actually do things instead of just chatting.

That’s it! You take an LLM, specialize it for a task, give it tools to take action and boom—you’ve got an agent. It’s really that simple! ( Even Perplexity AI can be thought of as a very advanced Google Search Agent! )

A Typical Agent Framework

Before we move forward to explore how agents can make our lives better and simpler, it's crucial to first understand the fundamental building blocks of an agent and how it is structured.

The image I have used here depicts an Agent framework probably in the easiest way possible.

In the image, the BRAIN icon denotes the LLM in use (essentially the core functioning unit of any Agent). The other three components connected to the Agent can be called its attributes.

TOOLS (API Access & Permissions)

These are the external functions the agent can use. This could be API access, database queries, or other system integrations that help the agent fetch real-time information or perform specific tasks.

For example, Gmail API access can be given to an agent that sends emails.

MEMORY (Docs, PDFs, Databases)

Memory allows the agent to store and retrieve information. It could be past interactions, stored documents, or any structured data that helps provide better responses.

Without memory, the agent would treat every request as new and have no context from previous interactions.

PLANNING (Base Prompt / System Prompt)

Planning is the predefined instruction that tells the agent how to behave and respond. Think of it as the agent's guidelines or personality settings. A well-structured prompt ensures that the agent stays on track and provides reliable, relevant responses.

For example, if you are creating a customer support agent, the planning part might include a prompt like:

"You are a customer support assistant for an e-commerce store. Your job is to provide polite, helpful, and clear answers to customer inquiries. Always maintain a professional yet friendly tone. If you don’t know the answer, suggest the customer contact support via email."

Multi-Agent Systems (MAS)

I’m sure you’ve noticed that whenever people talk about AI agents, the focus is almost always on multi-agent systems rather than single-agent ones. Why? Because a single agent, while powerful, comes with its own set of limitations. It can execute tasks, sure—but it works in isolation. It lacks specialization, scalability, and the ability to handle complex workflows efficiently, and that is exactly why the most important and helpful thing about AI Agents is Multi-Agent Communication Systems. Let's Understand this using an example -

Say the task is to automate the design and delivery of Email Campaigns.

The Single Agent Approach:

A flowchart illustrating a single-agent email automation system. The process starts with a user prompt, which is processed by an Email Agent using ChatGPT-4. The agent is responsible for generating the entire email, including the subject line, body content, and images. It has access to memory (email database) and uses the Gmail API to send the final email campaign. The diagram highlights how a single agent is burdened with multiple creative and analytical tasks. — "Diagram illustrating a single agent approach for automating email campaigns, powered by ChatGPT-4, handling tasks from generating subject lines to composing content and image descriptions."

If you look at the image, you’ll see how the single-agent system operates. The Email Agent (center) is responsible for handling the entire email workflow—right from generating the subject line to crafting the body content and even selecting images. While this setup seems convenient, the problem arises when a single agent is overloaded with multiple creative and analytical tasks (as seen in the image). This not only reduces efficiency but also affects reliability.

Now, consider the main objective behind AI agents—achieving complete autonomy. For an AI system to function independently, you need to trust it to handle tasks without constant supervision. However, in a single-agent setup, where one agent is burdened with too many responsibilities, its performance becomes unpredictable. The more tasks it juggles, the harder it is to trust it for full autonomy.

A detailed workflow diagram of a multi-agent email automation system, showcasing multiple specialized agents working together. The Campaign Manager delegates tasks to three specialized agents: Content Expert (Claude 3) for email body content, Subject Line Expert (ChatGPT-4) for subject line optimization, and Graphic Designer (DALL·E) for visual assets. Their outputs are collected and compiled by the Compiler Agent, which formats and sends the final email. This diagram highlights the advantages of multi-agent systems in handling complex tasks more efficiently than a single-agent approach. — Multi-Agent Framework for Email Campaigns: A diagram illustrating the collaboration among specialized agents, including a Campaign Manager, Content Expert, Subject Line Expert, Graphic Designer, and Compiler Agent, to efficiently create and execute an email campaign.

And here is the multi-agent framework, a structured approach that distributes tasks across specialized AI agents, mimicking how a human marketing team operates. Unlike a single-agent system, where one model is overloaded with multiple responsibilities, this workflow showcases clear role-based specialization, making the system more efficient and scalable. Each agent in this framework has a specific role, just like a well-organized marketing team:

The Campaign Manager acts as the team leader, delegating tasks efficiently.
The Content Expert focuses solely on writing persuasive email copy.
The Subject Line Expert specializes in crafting high-converting subject lines.
The Graphic Designer ensures visually appealing creatives.
The Compiler Agent assembles everything into a finalized email and sends it.

This division of labor ensures that no single agent is overwhelmed, allowing for better accuracy, faster execution, and improved decision-making. Just like in real teams, where a writer doesn't design images and a strategist doesn’t write copy, AI agents excel when given focused, specialized tasks.

Another major advantage of multi-agent systems is that we are not limited to a single LLM for all tasks. Instead, we can assign different LLMs based on their strengths

(We’ll take a deeper look into what goes into choosing the right LLM in the next blog—Design.)

You see, this workflow is just a sneak peek into how multi-agent systems work, it's probably the simplest one out there! In reality, they can get way more sophisticated and powerful, with extra agents, smarter feedback loops, and even real-time data crunching to fine-tune decisions. The possibilities? Endless. It's all up to your creativity!

And that’s exactly why multi-agent systems are the way forward. They’re faster, smarter, and way more efficient than overloading a single AI with everything. This isn’t just an improvement—it’s the future.

Conclusion

And that’s a wrap on Dream, but we’re just getting started. We’ve only scratched the surface of what’s possible with AI agents, and there’s so much more to explore. This was just the big picture, but in the next blog—Design—we’ll dive deeper into different types of multi-agent frameworks, breaking down how they work and what goes into building an effective system. We’ll also explore key considerations for designing a multi-agent setup.

If this felt like an eye-opener, just wait until we start building these systems step by step. Stay tuned—things are about to get even more interesting. Meet you in the next blog - Design!