Key Concepts
- Agentic: Using nondeterminism in a generative AI LLM to perform a task or set of tasks where the specific outcome is not the same every time.
- Agentic Workflows: More deterministic and focused on predictable decision-making and outcomes.
- AI Agents: Given agency to make decisions to accomplish a goal or outcome autonomously.
- Nondeterminism: The characteristic of an agent where the outcome of a task is not the same every time.
- Autonomy: The ability of an agent to perform tasks and make decisions without direct human intervention.
- Reflection: The process where an agent evaluates its progress and adjusts its actions to better achieve the desired outcome.
Defining Agentic Behavior
Jason Davenport defines "agentic" as using nondeterminism in a generative AI LLM to perform a task or set of tasks where the specific outcome is not deterministic or the same every time. Agentic behaviors can be combined into workflows or AI agents to drive towards a specific outcome.
AI Agents vs. Agentic Workflows
Aja Hammerly distinguishes between AI agents and agentic workflows. Workflows are more deterministic, focusing on predictable decision-making. AI agents, on the other hand, are given agency to make decisions to achieve a goal autonomously. Many tasks are better suited for agentic workflows than full AI agents. A true AI agent can pick any number of tasks in any order to accomplish the outcome, whereas a workflow will be more probabilistic and predictable.
Examples of Workflows and Agents
Jason provides an example of a workflow for processing invoices. The workflow uses an LLM with tools to analyze the invoice, scrape data, and input it into a form. The user reviews the information and uploads it to a processing tool.
An example of an agent is a software agent tasked with creating a website for the game "Snake." The agent defines and performs tasks, reflecting on its progress to ensure the outcome meets the specified goal. This is all done autonomously.
Importance and Application of Agentic Behavior
Jason emphasizes the importance of understanding how much agency is appropriate in different situations, drawing parallels to human interaction in business settings.
Aja notes that many current problems are better solved with agentic workflows than full AI agents. Adding a single or a couple of LLM steps into an existing deterministic workflow (e.g., for processing or summarization) can provide significant benefits.
Considerations for Building Agentic Systems
When building agentic flows, it's crucial to consider the risks, both positive and negative, and the potential consequences of making the system fully agentic. For example, a fully agentic customer service agent might make decisions a human wouldn't. In such cases, scaling back to a workflow with human-in-the-loop may be preferable.
Agents are best suited for situations where a goal and outcome can be defined, a logical trigger for the agent exists, and autonomy is desirable.
Practical Resources and Examples
Jason mentions that code snippets and architectural examples will be published on GitHub to help viewers implement the concepts discussed. These resources aim to make abstract topics more implementable for specific use cases.
Conclusion
The discussion emphasizes the importance of understanding the nuances of agentic behavior and the distinction between agentic workflows and AI agents. It highlights the need to carefully consider the risks and benefits of autonomy when designing AI systems and to choose the appropriate level of agency based on the specific task and desired outcome. The upcoming code examples and resources promise to provide practical guidance for implementing these concepts.
AI summaries can miss context or contain errors. Check important details against the original video.