Summary
Highlights
This video is the second in a series on mastering AI Agent Engineering. It will provide a deeper theoretical understanding of AI Agents, drawing on a paper from Anthropic. The theoretical knowledge is crucial for interviews and for understanding the growing field of AI Agent Engineering.
An AI Agent can be defined as a digital worker, an AI-powered entity that can sense its environment and make decisions to achieve its goals. This definition aligns with the previous video's example of a car with a driver (AI Agent) using tools (car controls) to navigate. It emphasizes that an AI Agent combines an LLM with a set of tools.
Another definition describes AI Agents as programs where the output of the Large Language Model (LLM) controls the workflow. Unlike traditional LLMs that only answer questions, an AI Agent can interact with its environment and take actions, such as sending emails automatically, which an LLM alone cannot do.
An 'agentic system' can be identified by several features: its ability to call LLMs, its capacity to interact with various tools (like Gmail), its capability to interact with its environment (like autonomous cars using sensors to make decisions), its ability to plan and prioritize tasks, and finally, its autonomy to make decisions without human intervention.
Anthropic's paper distinguishes between two types of agentic systems: Workflows and AI Agents. Workflows use LLMs and tools but follow a predefined path, with limited decision-making by the LLM (around 10%). In contrast, AI Agents leverage LLMs and tools to make dynamic, real-time decisions (up to 80%) to achieve a goal, adapting to the environment and not following a fixed process.
Understanding the difference between workflows and true AI Agents is crucial, as many systems currently labeled as AI Agents are, in reality, workflows due to their predetermined actions rather than autonomous decision-making. True AI Agents, like autonomous vehicles, exhibit high levels of autonomy.
Anthropic highlights that both workflows and agents are agentic systems, but a fine line differentiates them: the level of automation and autonomous decision-making. Workflows have low automation (around 10% LLM-driven decisions), while true AI Agents have high automation (over 80% LLM-driven decisions), allowing the system to think and make choices to reach its goal.