“Agents are not only going to change how everyone interacts with computers. They’re also going to upend the software industry, bringing about the biggest revolution in computing since we went from typing commands to tapping on icons.” — Bill Gates, Co-founder of Microsoft
“As agents become more widespread more intelligent and more sophisticated, it’ll likely change the way we think about computers in the first place — in the same way that the transition from a command line interface to a graphical interface completely revolutionized the way we interact with computers.” — Daoud Abdel Hadi, TEDxPSUT Speaker
“AI agents will become the primary way we interact with computers in the future. They will be able to understand our needs and preferences, and proactively help us with tasks and decision making.” — Satya Nadella, CEO of Microsoft
“By 2024, AI will power 60% of personal device interactions, with Gen Z adopting AI agents as their preferred method of interaction.” — Sundar Pichai, CEO of Google
“AI agents will become our digital assistants, helping us navigate the complexities of the modern world. They will make our lives easier and more efficient.” — Jeff Bezos, Founder and CEO of Amazon
“We could only be a few years, maybe a decade away [from general artificial intelligence].” — Demis Hassabis, Co-founder and CEO of DeepMind
“AI agents will transform the way we interact with technology, making it more natural and intuitive. They will enable us to have more meaningful and productive interactions with computers.” — Fei-Fei Li, Professor of Computer Science at Stanford University
“AI agents will become an integral part of our daily lives, helping us with everything from scheduling appointments to managing our finances. They will make our lives more convenient and efficient.” — Andrew Ng, Co-founder of Google Brain and Coursera
“I don’t think we’ve kind of nailed the right way to interact with these agent applications. I think a human in the loop is kind of still necessary because they’re not super reliable.” — Harrison Chase, Founder of LangChain
“For a long time, we’ve been working towards a universal AI agent that can be truly helpful in everyday life.” — Demis Hassabis, Co-founder and CEO of DeepMind
2025 is shaping up to be the year of AI agents. But to understand what that means, you need to look at the changes happening in AI.
From Monolithic to Compound AI
The first big shift is moving from monolithic models to compound AI systems. Monolithic models are trained on a fixed dataset, and that limits what they know and what they can do. They’re hard to adapt. If you want a model to perform better, you need more data, more time, and more resources.
Think of asking a monolithic model how many vacation days you have left. It might generate a response, but it’s not going to be right. The model doesn’t know who you are or have access to your HR database. It’s working with what it was trained on, not real-time data.
That’s where compound AI comes in. A compound AI system takes a model and integrates it with other components, like databases, APIs, or verification tools. Instead of relying solely on the model, the system can search a database, retrieve the exact number of vacation days left, and give you a correct answer.
Compound AI is modular. You combine the right tools for the task. You don’t need to retrain a massive model; you can plug in different components as needed. This makes compound systems easier to build and adapt.
System Design in AI
Designing a compound AI system is like system design in software. You don’t build everything from scratch. You break down the problem, pick the right components, and connect them. The model is just one part. You might add tools to search databases, verify outputs, or do calculations. The result is a system that’s flexible and faster to adjust than retraining a giant model.
Retrieval-augmented generation (RAG) is a popular type of compound AI. RAG systems use a model to generate responses, but they also search for real-time information. If your system only searches your HR database, it works great for vacation queries but fails if you ask about the weather. The system’s logic is fixed. It follows a set path.
Enter AI Agents
AI agents solve this limitation by putting a large language model (LLM) in charge of the system’s logic. The recent improvement in LLM reasoning capabilities makes this possible. Instead of following a fixed path, agents can plan and adjust their approach dynamically.
An agent doesn’t just generate an answer. It thinks through the problem. It breaks it down, plans each step, and adjusts if things go wrong. It’s the difference between getting an instant answer and taking time to solve the problem thoughtfully.
What Makes an AI Agent
Agents have three core capabilities: reasoning, acting, and memory.
1. Reasoning: The model plans how to solve a problem. It doesn’t just respond; it thinks through each step.
2. Acting: The agent uses tools to execute the plan. These tools might include:
- Searching the web
- Querying databases
- Doing calculations
- Calling other AI models (e.g., for translation or summarization)
Tools extend what the model can do — and the agent decides when and how to use them.
3. Memory: Agents can remember previous steps or conversations. This makes interactions more personalized. Memory can be logs of the agent’s thought process or a history of your interactions with it.
The React Framework
A popular way to build agents is the React framework, which combines reasoning and acting. When you ask a React agent a question, it doesn’t just generate an answer. It plans, acts by using tools, and adjusts until it gets the right solution.
Let’s go back to the vacation example. Imagine you want to know how many sunscreen bottles you need for a trip to Florida. That’s a complex problem. An agent might:
- Check your vacation days.
- Look up the weather forecast.
- Find the recommended sunscreen dosage.
- Do the math to calculate how many bottles you need.
A simple model wouldn’t handle this. But an agent can reason through the steps, use tools to gather information, and get the right answer.
The Future of AI Systems
Compound AI systems are here to stay. What’s changing is how much autonomy we give them. For simple, well-defined tasks, fixed logic is more efficient. For complex tasks with many unknowns, agents are better. They can explore different paths and adjust as needed.
We’re in the early days of AI agents. But the combination of system design and agent reasoning is powerful. The more agents can plan and adapt, the more complex tasks they’ll handle. For now, a human will often stay in the loop. But as agents improve, they’ll take on more autonomy.
In 2025, we’re going to see more AI agents integrated into systems. And as they get better at reasoning and acting, they’ll become even more useful.
We’re not just building smarter models. We’re building smarter systems.
FAQ: AI Agents and Compound Systems
What are AI agents?
AI agents are systems that use large language models (LLMs) to dynamically reason, plan, and execute tasks by integrating external tools and memory.
What is a compound AI system?
A compound AI system combines a core model with other components like databases, APIs, and verification tools to enhance flexibility and accuracy.
How does the React framework work?
The React framework allows AI agents to plan steps, use tools, and adjust dynamically until they solve a problem correctly.
What is the difference between monolithic AI and compound AI?
Monolithic AI relies on a fixed dataset, while compound AI integrates various tools and data sources for real-time problem-solving.
Why are AI agents important for the future?
AI agents offer more adaptability and autonomy, making them suitable for solving complex, multi-step tasks.
What is the difference between a monolithic AI model and a compound AI system?
A monolithic AI model is trained on a fixed dataset, limiting its knowledge and adaptability. It relies solely on its training data, making it difficult to provide accurate answers to questions requiring real-time information or access to external data. Compound AI systems, in contrast, integrate a model with other components like databases, APIs, and verification tools. This allows them to access real-time information and perform tasks that a monolithic model alone could not handle, making them more flexible and adaptable.
How does “system design” apply to creating AI systems, and what does this entail?
System design in AI involves breaking down a problem into smaller components and selecting appropriate tools and models to address each part. Instead of building everything from scratch or retraining a massive model, AI systems are designed to use existing components in a modular fashion. This approach often involves integrating tools for database searches, output verification, and computations to create a system that is more versatile and easier to adjust. The core AI model is only one component of the overall system.
What is Retrieval-Augmented Generation (RAG), and what are its limitations?
Retrieval-Augmented Generation (RAG) is a type of compound AI that uses a model to generate responses but also searches for real-time information to supplement its knowledge. While RAG systems are effective at accessing and integrating information from defined sources (e.g., a company’s HR database), their logic is fixed to a set path. This means they are limited in their ability to handle questions requiring information from outside the sources it is configured to use. For example, if only connected to HR databases, a RAG model would not be able to answer questions about the weather.
What distinguishes AI agents from other AI systems, and what core capabilities do they possess?
AI agents differ from other AI systems by having the ability to dynamically plan and adjust their approach to solve problems rather than following a pre-defined path like RAG. They have three core capabilities: reasoning, acting, and memory. Reasoning involves planning how to solve a problem step-by-step. Acting is about using external tools to execute the plan, such as searching the web or querying databases. Memory allows them to remember previous steps and interactions, leading to more personalized experiences.
How does the “acting” capability of AI agents work and what are examples of the tools they utilize?
The “acting” capability of AI agents allows them to execute plans by utilizing external tools. These tools are external programs that expand what the agent can do, such as searching the web, querying databases, performing calculations, or using other AI models like translation models. The core model determines when and how to utilize the tool in order to solve the current problem, allowing the agent to go beyond just generating a response, and act upon it.
What is the React framework, and how does it work in practice?
The React framework (Reasoning and Acting) is a method for building AI agents where the agent doesn’t just respond but actively plans, acts by using various tools, and adjusts its approach until it achieves the desired result. When presented with a complex problem, a React agent will go through a thought process that includes breaking the problem down into manageable parts, then utilizing the necessary tools for data gathering and performing the necessary steps to achieve a correct solution.
Why are AI agents better suited for complex tasks compared to fixed-logic systems?
Fixed-logic systems are more efficient for simple, well-defined tasks, where the path to the solution is clear. However, for complex tasks with many unknowns, AI agents are better because they can explore multiple paths and dynamically adjust their approach based on what they learn. They can engage in a thought process, planning, then taking action, which allows them to solve problems in a dynamic, much more effective manner compared to static systems.
What is the predicted future direction of AI agents, and what role will humans play in their development and use?
The future direction of AI agents involves increased integration into various systems and a higher degree of autonomy as their reasoning and acting capabilities improve. While AI agents are becoming more capable, human oversight will likely remain an important part of the process, at least initially. Humans will help shape the development and ensure proper usage of these complex and powerful systems. As the agents advance, they will be able to handle more complex and undefined tasks, but the goal is not for them to replace human oversight entirely, but to create more effective AI systems.
Top comments (0)