Andrew Ng on Agentic Workflows

I recently saw a video by Andrew Ng on Agentic workflows, which was fairly enlightening, especially if you are looking to learn more about LLM-based Agents. You can watch the full video at the link below. Following is a short summary of the video.

A New Way to Interact with Large Language Models

Agentic workflows fundamentally change how we interact with LLMs by making it an iterative process. Instead of a one-shot prompt, users and LLMs engage in a back-and-forth conversation. The user provides an initial prompt, the LLM clarifies or asks questions, and then refines its response based on this conversation. This collaborative approach goes beyond simply generating text — it allows for a more nuanced and effective use of LLMs.

Agentic Workflows: A Paradigm Shift

Agentic workflows unlock a new range of capabilities for LLMs. Here are a few key design patterns Ng proposes:

• LLMs as Code Reviewers: Imagine an LLM that can analyze its own code, find bugs and suggest fixes. This would be a game-changer for programmers!
• Planning and Teamwork: LLMs can be used to plan complex tasks and collaborate with other AI agents to complete them. Ng describes a system where one LLM acts as a coder, writing code, while another acts as a reviewer, checking for errors.
• LLMs and Other Tools: Agentic workflows can combine LLMs with specialized tools to create even more powerful workflows. For example, an LLM could write code and another tool could compile and run it.
• Speed Over Perfection: Because agentic workflows involve frequent back-and-forth communication, it’s important for LLMs to prioritize speed over absolute perfection in their initial responses. This allows for faster iteration and refinement.

Conclusion: The Future of LLMs

Agentic workflows represent a significant step forward in unlocking the full potential of LLMs. Ng believes this approach is a stepping stone towards achieving artificial general intelligence (AGI). However, it also requires a change in mindset. We need to move away from expecting an immediate answer and embrace the collaborative nature of agentic workflows.

Overall, agentic workflows offer a promising new way to interact with LLMs, opening doors to a future of more powerful and versatile AI tools.