What are AI Agents?

Key Concepts

• Monolithic Models vs. Compound AI Systems: The shift from single, large models to systems composed of multiple components.
• Compound AI Systems: AI systems built with modular components, including models, programmatic elements, databases, and tools.
• Retrieval Augmented Generation (RAG): A popular type of compound AI system that retrieves information to augment the generation process.
• Control Logic: The pre-defined path a compound AI system follows to answer a query.
• AI Agents: Compound AI systems where a Large Language Model (LLM) controls the logic and decision-making.
• Reasoning: The ability of an LLM to break down complex problems and create a plan to solve them.
• Tools: External programs or APIs that an AI agent can use to perform specific actions.
• Memory: The ability of an AI agent to store and retrieve information, including internal logs and conversation history.
• ReACT: A popular framework for configuring AI agents that combines reasoning and acting.
• AI Autonomy: The degree to which an AI system can operate independently.

1. The Shift from Monolithic Models to Compound AI Systems

• Limitation of Monolithic Models: Models are limited by their training data, impacting their knowledge and task-solving abilities. They are also difficult and resource-intensive to adapt.
• Example: Vacation Planning: A monolithic model cannot accurately determine the number of vacation days available to a specific user because it lacks access to personal data.
• Compound AI Systems: Integrate models into existing processes, giving them access to external data and tools.
• Example: Vacation Planning (Compound System): The query is fed into a language model, which generates a search query for a database containing vacation data. The database returns the information, which is then used by the model to generate a correct answer.
• Modularity: Compound AI systems are modular, allowing for the selection of appropriate components (models, programmatic elements, output verifiers, databases, tools) to solve specific problems.
• Adaptability: Compound AI systems are faster and easier to adapt than tuning a monolithic model.

2. Compound AI Systems and Control Logic

• Retrieval Augmented Generation (RAG): A common type of compound AI system.
• Limitations of Predefined Control Logic: Most RAG systems have a predefined control logic, meaning they follow a specific path to answer queries. This can lead to failures if the query falls outside the expected scope.
• Example: Weather Query: A RAG system designed to access a vacation policy database will fail when asked about the weather.

3. The Emergence of AI Agents

• LLMs in Control: AI agents use LLMs to control the logic of a compound AI system. This is possible due to advancements in LLM reasoning capabilities.
• Think Slow vs. Think Fast: Systems can be designed to "think fast" (act as programmed) or "think slow" (create a plan, break down the problem, and adjust the plan as needed).
• Agentic Approach: Putting an LLM in charge of the logic is considered an agentic approach.

4. Components of LLM Agents

• Reasoning: The LLM is at the core of problem-solving, creating plans and reasoning about each step.
• Acting (Tools): Agents use external programs or APIs (tools) to perform actions.
  • Examples of Tools: Web search, database search, calculators, code execution, translation models, APIs.
• Memory: Agents can access memory to store and retrieve information.
  • Types of Memory: Internal logs of the model's reasoning process, history of conversations with the user.

5. ReACT Framework

• ReACT (Reasoning and Acting): A popular framework for configuring AI agents.
• Process:
  1. User Query: The user submits a query.
  2. Prompting the LLM: The LLM is given a prompt to think slow and plan its work.
  3. Acting: The LLM decides whether to use external tools.
  4. Tool Call (if needed): The LLM calls a tool and receives an answer.
  5. Observation: The LLM observes the answer and determines if it addresses the query.
  6. Iteration (if needed): The LLM iterates on the plan and tackles the problem differently until a final answer is reached.
• Example: Sunscreen Calculation:
  • Query: "I'm planning to go to Florida next month and will be outdoors a lot. I'm prone to burning. How many two-ounce sunscreen bottles should I bring?"
  • Reasoning and Planning: The agent needs to determine:
    • How many vacation days are planned (potentially retrieved from memory).
    • How many hours will be spent in the sun (requires weather forecast lookup).
    • Recommended sunscreen dosage per hour (requires accessing a public health website).
    • Perform calculations to determine the number of two-ounce bottles needed.

6. AI Autonomy and the Future of AI Agents

• Sliding Scale of AI Autonomy: The level of autonomy in an AI system can be adjusted based on the specific problem.
• Trade-offs:
  • Narrow, Well-Defined Problems: A programmatic approach (predefined control logic) can be more efficient than an agentic approach.
  • Complex Tasks with a Wide Range of Queries: An agentic approach is more suitable because it is too difficult to configure every possible path in the system.
• Early Days of Agent Systems: The field is rapidly evolving, combining system design with agentic behavior.
• Human in the Loop: Humans will likely remain in the loop to ensure accuracy.

7. Synthesis/Conclusion

2024 is predicted to be the year of AI agents, marking a significant shift from monolithic models to more adaptable and intelligent compound AI systems. These systems, particularly those leveraging LLMs for control logic (AI agents), offer enhanced reasoning, access to external tools, and memory capabilities. Frameworks like ReACT enable agents to tackle complex problems through iterative planning and execution. While programmatic approaches remain efficient for narrow tasks, agentic systems excel in handling complex, varied queries. The future of AI lies in carefully balancing AI autonomy with human oversight, paving the way for more sophisticated and versatile AI solutions.