AI Agents: The Future of AI and Its Potential

AI Agents, the next frontier in AI, hold immense potential to drive significant progress. They build upon the surprising advancements of prompt engineering, extracting the utmost value and reasoning from any LLM model. These agents represent a leap forward in our ability to reason about arguments, solve complex problems, make informed decisions, and assess the impact of these decisions. 

We have mostly been using LLMs in zero-shot mode, hoping to get the best outcome with a single prompt. It is like writing an essay without a chance to review and critique it. Despite the limited chances we give the LLM with this mode, most notable LLMs out there are doing amazingly well. 

Andrew Ng and his research team tested and compared zero-shot with AI agents to achieve certain tasks. As you see in the diagram below, AI Agents significantly improved the performance of the LLM models. This made ChatGPT 3.5 very close in performance to ChatGPT 4.0 despite the enormous difference in model size. 

We already read the news about Devin, your AI-based software engineer, and how the impressive promise it carries. Devin is a group of agents working with different perspectives and capabilities to solve your application development task!

Why am I excited to write about AI agents?

AI agents are not just tools, they are the catalysts for a transformative shift in AI. They endow LLMs with superpowers, enabling them to take generated outputs and perform necessary tasks or further analyze the job. This is the ultimate promise of building advanced AI systems. We’ve been building passive tools and automation for many solutions, but these tools lack the ability to resonate when we encounter a problem or have a goal to achieve. AI agents are a significant step towards mimicking our thought process, with LLMs’ ability to understand and resonate about diverse concepts and topics. 

As an entrepreneur and a startup enthusiast, I understand the challenges of innovating with LLMs. The resources required to develop your own specialized LLM are often only available to the existing big players or AI celebrities who have attracted enough capital and resources. However, with AI Agents, you can develop a deep enough IP on top of LLMs without getting entangled in the complex economics of building your LLMs, at least for now. This accessibility is a game-changer. 

Agents open the door to solving serious problems. In a later example, I discuss an agent-based system that looks for application vulnerabilities and proactively secures them. LLMs become the underlying compute layer that disappears in the background. Think of how IaaS was the star in the early days of cloud computing. Then, Platform as Service and serverless workloads became more common technologies to build exciting applications. 

Are AI Agents our way to Artificial General Intelligence (AGI)?

AI agents are a step towards AGI. In simple terms, agents act like our inner voice or smart virtual assistants, analyzing problems and bouncing ideas to reach conclusions and decisions. In addition, they can execute these ideas in their virtual world. They are not simple reflex agents, though. They are not simple condition action rules we used to see in old AI agents. AGI is the final result of having a group of agents working together on a specific problem that hopefully surpasses human performance in solving the same problem. 

AGI is a field of artificial intelligence and the holy grail of AI. Researchers have been working on different AGI models to allow machine learning to work independently on other tasks, including learning new skills. The last two years revived that concept and made it possible to get sensible results out of basic agentic workflows, such as the famous BabyAGI project that made a lot of splash last year. 

AI Agents vs Prompt Engineering

You may argue that AI agents are just a glorified way to build a workflow around prompts and the outputs they generate. In some limited use cases of AI Agents, yes, AI agents could be another way to organize prompts and create simple workflows around them. But when thinking about an outcome generated by an LLM model, using another model to refine the answer, and validating the result in a different system, this becomes an interesting case for agents and builds productive agentic workflows. Think of writing a cloud service that scans a repo for security vulnerabilities. This is a perfect case where you want developers and security engineers to constantly update the application, including new scans and strategies to scan and fix vulnerabilities. Imagine you have a group of agents collaborating on building the service and updating it in the background based on what is discovered and reported on the public Internet and the underground. That becomes very powerful quickly, primarily if another agent reports application updates in each release. Agents, including PMs, can become your entire software development team to develop such a critical service. Prompt engineering becomes a very tiny piece in the whole process. 

How Can Agents Work Together? 

Open-source AI agents are expanding very rapidly, which makes it exciting but challenging at the same time. Categorizing different types of agents and how they can interact together might help us better understand what’s out there and hopefully determine the proper system to use in your project or product idea. You are probably thinking by now, how can I build my agent framework 😉 

I like very much Andrew Ng’s taxonomy of agent workflow patterns. These patterns tell you that agents can work together to achieve a goal. Remember, agents are supposed to mimic our way of thinking, which can go in many ways. Below is just a quick overview of each category. However, I plan to dive deeply into each category to discover the pros and cons of each one. 

• Reflection Agents: These agents work together to review, examine, and critique generated LLM outputs. Agents can depend on the same LLM or have their own model. The LLM examines its own work to find ways to improve it. 
• Tool use Agents: In this category, agents go beyond the world of the LLM to perform tasks, validate results, or use them to perform the next step in the given task. For example, they compile and test LLM code generated by the LLM model. 
• Planning Agents: Agents in this pattern work on setting goals, sub-goals, and so on to create a complete executable plan out of a general goal given to them.
• Multi-agent collaboration: In this multi-agent systems pattern, agents collaborate to work on a task or debate an idea. This is similar to the brainstorming exercise that humans typically do in problem-solving situations.

Concluding Remarks

AI research and open-source communities are excited about agents because of their potential to significantly improve LLMs. Agents open the door to building specialized systems on top of already impressive LLMs and solving real problems. Understanding their different patterns, current limitations, and when we can use them is key. In the next set of articles, I’ll discuss some key points defining agentic workflow patterns and frameworks and when you should use each.