An Introduction to AI Agents (for 2025)

Summary

In this video, Shaw introduces the concept of AI agents, exploring their definition, features, and potential. With 2025 marked as the year of AI agents, Shaw discusses the central role of large language models (LLMs), tool use, and autonomy in these systems. Contrasting AI agents with traditional chatbots, Shaw highlights the value of tools in bridging the gap between LLMs and the real world. The video also outlines three levels of agency in AI systems, from basic LLM-plus-tools to fully autonomous agents. Shaw emphasizes the growing potential of AI agents and invites viewers to contribute ideas for future videos in the series.

Takeaways

• 😀 AI agents are systems that utilize large language models (LLMs) along with tools to perform tasks, but there's no single agreed-upon definition for what an AI agent actually is.
• 😀 OpenAI, Hugging Face, and Anthropic each provide slightly different definitions of AI agents, with a focus on LLMs, tools, planning, and autonomy.
• 😀 A key feature of AI agents is the integration of LLMs with tools like web search, code interpreters, and APIs to interact with real-world data and perform tasks beyond text generation.
• 😀 AI agents can exhibit different levels of agency, ranging from simple tool use (Level 1) to complex workflows (Level 2) and autonomous loops (Level 3).
• 😀 Unlike traditional chatbots, AI agents are more capable because they can interact with reality through tools, taking real-world feedback to improve their performance.
• 😀 One important advancement in AI agents is the concept of 'test time compute scaling,' which allows models to reason and plan more effectively before providing answers.
• 😀 While agents can solve tasks with less detailed instructions due to their ability to reason, their value lies in how they use real-world feedback through tools like code interpreters and memory.
• 😀 LLM-based tools can address major blind spots of traditional LLMs, such as the inability to access up-to-date information, run code, or interact with dynamic environments.
• 😀 Level 2 agentic systems, called 'LLM workflows,' use predefined steps to solve more complex tasks, utilizing multiple LLMs and tools working together.
• 😀 At Level 3, AI agents operate in a feedback loop, where the LLM generates a response, is evaluated, and receives feedback to continuously refine its output until it meets specified criteria.

Q & A

What is an AI agent, and why is it important?

An AI agent is typically defined as an LLM (large language model) combined with tools and instructions to perform specific tasks autonomously or semi-autonomously. AI agents are important because they can interact with real-world data and perform more complex tasks, which traditional LLMs could not do alone, making them valuable for solving real-world problems.

Why is there confusion about the definition of AI agents?

The confusion stems from different organizations providing slightly different definitions. For example, OpenAI focuses on tools, Hugging Face emphasizes planning, and Anthropic stresses autonomy. These differing focuses on tools, planning, and control lead to a lack of consensus on a single definition of AI agents.

What are the three key features of AI agents?

The three key features are: 1) An LLM is involved, which is central to all AI agents. 2) Tool usage, which expands the LLM’s capabilities by allowing interaction with the real world. 3) Autonomy, where the agent has some level of control over how it accomplishes tasks, often involving reasoning or reflecting on outputs.

How do AI agents differ from traditional chatbots?

AI agents differ from traditional chatbots because they can interact with the real world through tools and feedback. Unlike chatbots that can only generate text, AI agents can use tools like web searches, code execution, and other APIs to gather real-world information or take actions, making them more capable of solving complex, real-world problems.

What role do tools play in enhancing the capabilities of AI agents?

Tools are crucial for AI agents because they allow LLMs to access and interact with the real world. Tools like web searches, Python interpreters, and API calls provide real-time data and actions that go beyond text generation, enabling the agent to perform tasks like running code, retrieving up-to-date information, or taking actions on behalf of the user.

What is 'testime compute scaling,' and how does it improve AI agent performance?

'Testime compute scaling' refers to the idea that giving an LLM more time to generate responses improves its performance. By allowing the model to plan and think through tasks before generating an answer, the AI agent can provide more accurate and complex responses without requiring specific instructions for every task.

What are LLM workflows, and how do they contribute to AI agent development?

LLM workflows are predefined sequences of steps that involve at least one LLM. These workflows enable more reliable and complex task completion by breaking down tasks into simpler steps. Instead of relying on a single model to complete a task, workflows allow multiple LLMs to collaborate and address different parts of a problem, improving efficiency and performance.

Can you explain the difference between level one and level two AI agents?

Level one agents are simply LLMs combined with tools, providing basic interaction with the real world. Level two agents, on the other hand, involve LLM workflows, where multiple LLMs and tools are orchestrated to handle more complex tasks. Level two systems can handle tasks that require multiple steps or decisions, improving reliability and performance.

What is the role of parallelization in LLM workflows?

Parallelization in LLM workflows involves running multiple tasks simultaneously to improve speed and reduce latency. There are two types: sectioning, where tasks are split into subtasks that can run in parallel, and voting, where multiple LLMs perform the same task, and their results are combined to improve accuracy.

What is an LLM in a loop, and how does it enhance AI agent autonomy?

An LLM in a loop refers to a system where an LLM generates outputs, which are then evaluated and refined through multiple iterations until the desired outcome is achieved. This closed-loop process allows the AI agent to perform open-ended tasks autonomously, adapting and improving based on feedback, which makes it capable of handling more complex, real-world tasks without explicit instructions.