What are AI Agents?

Introduction

2024 is poised to be the year of AI agents. To understand AI agents, we need to examine the significant shifts in the field of generative AI. One noteworthy shift is the transition from monolithic models to compound AI systems.

From Monolithic Models to Compound AI Systems

Models alone have limitations due to the data they are trained on, which affects the tasks they can solve and their adaptability. For instance, tuning a model requires significant data and resources.

Example: Planning a Vacation

Suppose I want to plan a vacation and determine how many days of vacation I have left. Inputting this query into a standard model would yield an incorrect answer, as the model lacks access to my personal data. However, a compound AI system can integrate the model into existing processes, like accessing a database containing my vacation information.

Designing a Compound AI System

To address this problem:

Query: Input my vacation query into the language model.
Database Search: The model generates a search query for the vacation database.
Response Generation: Retrieve the information, and the model generates a correct sentence response.

This compound AI system demonstrates that certain problems are better solved using principles of system design, involving multiple modular components like tuned models, image generation models, output verifiers, and searchable databases.

The Role of AI Agents

AI agents control the logic of compound AI systems by leveraging advancements in large language models' reasoning capabilities. These agents can reason, act using external tools, and access memory.

Reasoning Capability

Large language models can break down complex problems and create plans, adjusting as needed to achieve success.

Acting Capability

Agents determine when and how to call external tools like web search, calculators, code, or translation models to optimize their problem-solving approach.

Memory Capability

Agents store inner dialogues and conversation history to enhance personalization and improve the interaction experience.

Configuring AI Agents: The REACT Approach

By combining reasoning and acting components, REACT agents can:

User Query: Input a query into an LLM with instructions to plan before responding.
Tool Utilization: Define when to use external tools to address the query.
Iteration: Adjust the plan and iterate until arriving at a final answer.

Example: Vacation Planning with Sunscreen

For a complex task like determining how many 2 oz sunscreen bottles to bring to Florida, an agent would:

Retrieve Vacation Days: Access memory for vacation days.
Check Weather Forecast: Look up the average sun hours in Florida.
Dosage Information: Research public health recommendations for sunscreen dosage.
Calculation: Calculate the total sunscreen needed in 2 oz bottles.

Compound AI Systems: The Future

As compound AI systems become more agentic, system designers can choose between programmatic routes for narrow, well-defined problems and agentic approaches for complex, diverse tasks. This balance can optimize efficiency and adaptability.

Conclusion

AI agents represent a significant advancement in AI, with the potential to handle increasingly complex tasks autonomously. As we continue to explore these technologies, their integration into everyday processes will become more seamless and efficient.

Keywords

AI agents
Compound AI systems
Large language models
Reasoning
Acting
Memory
REACT agents
Generative AI
Programmatic systems
Personalization

FAQ

What are AI agents?

AI agents are sophisticated AI systems that leverage reasoning, acting, and memory capabilities to autonomously solve complex tasks.

How do AI agents differ from traditional AI models?

Unlike traditional monolithic models that rely heavily on pre-trained data, AI agents use multiple modular components and external tools to solve problems, making them more adaptable and capable.

What is a compound AI system?

A compound AI system integrates various AI models and programmatic components, allowing it to perform more complex and personalized tasks compared to monolithic models.

What capabilities do AI agents possess?

AI agents can reason (create and adjust plans), act (use external tools for problem-solving), and access memory (store and retrieve inner dialogues and conversation history).

How is the REACT approach used to configure AI agents?

The REACT approach combines reasoning and acting components. It prompts the AI to plan its response and use external tools, iterating until it provides a final answer.