2024: The Year of AI Agents

What Are AI Agents?

To understand AI agents, we must first examine the shifts in the field of generative AI. The first significant shift is the transition from monolithic models to compound AI systems.

From Monolithic Models to Compound AI Systems

Models, when used in isolation, are confined by the data they are trained on, limiting their knowledge of the world and the tasks they can solve. Additionally, adapting these models can require substantial investment in data and resources.

Illustrating the Limitations of Traditional Models

For example, if I want to plan a vacation and ask a model how many vacation days I have, the response is likely to be incorrect since the model has no personal data about me. While models can assist with tasks like summarizing documents and drafting emails, they truly shine when integrated into systems that leverage their capabilities.

Building a Compound AI System

Imagine a system designed to answer my vacation query. By allowing the model to access the database storing my vacation information, it can generate a search query to retrieve accurate details. This interaction between multiple components exemplifies a compound AI system that effectively addresses complex tasks through system design principles.

Modularity in Systems

Compound AI systems are inherently modular, allowing for the integration of different models and programmatic components. This modularity simplifies problem-solving compared to merely tuning a model, enhancing adaptability and efficiency.

Retrieval-Augmented Generation (RAG)

One common type of compound AI system is Retrieval-Augmented Generation (RAG). However, RAG systems can become ineffective if they are not designed with varying queries in mind. Control logic, which is defined by the human operator, dictates how queries are processed.

The Role of AI Agents

Where do AI agents fit into this landscape? An alternative way to control the logic of a compound AI system is by placing a large language model (LLM) at the helm. Improvements in the reasoning capabilities of LLMs make this approach viable.

Characteristics of LLM Agents

LLM agents possess several key capabilities:

• Reasoning: The model can devise a plan and evaluate each step throughout the problem-solving process.
• Acting: Agents can utilize external tools—like web searches and calculators—to implement their solutions.
• Memory Access: Agents can store and retrieve internal logs or conversation histories to enhance personalization.

Configuring Agents: The ReACT Framework

One popular method for configuring agents is through the ReACT framework, which integrates the reasoning and acting components of LLM agents.

Applying the ReACT Framework: A Vacation Example

Consider my query about the number of sunscreen bottles needed for my upcoming vacation in Florida. This complex problem requires multiple steps:

1. Retrieving how many vacation days I have.
2. Estimating sun exposure hours.
3. Determining recommended sunscreen dosage.
4. Calculating how much sunscreen fits into two-ounce bottles.

This scenario underscores the modularity and capability of a compound AI system designed to tackle complicated tasks.

The Future of Compound AI Systems

In 2024, we will see the continued evolution of compound AI systems toward more agent-centric designs. There is a sliding scale of AI autonomy, and the system’s designer must weigh the trade-offs associated with various degrees of autonomy.

Narrow vs. Complex Problems

For narrow, well-defined problems, a programmatic approach may prove more efficient, as one can tailor responses to specific queries. However, for complex tasks, an agentic approach allows for greater flexibility and adaptability.

Conclusion

We are in the early days of agent systems, with rapid advancements expected. Combining system design principles with generative capabilities will lead to improved accuracy and functionality. As these developments unfold, a human will often remain in the loop to ensure reliability. Thank you for exploring the future of AI agents with us, and please subscribe for more insights!