Using Agentic AI to create smarter solutions with multiple LLMs (step-by-step process)

Agentic AI: A Deep Dive

Key Concepts:

• Agentic AI
• Compound LLMs
• AI Agents
• Orchestrator Agent
• Prompt Engineering
• Tools as Agents
• Dynamic Workflows

Compound LLMs: Enhancing AI Output

The video begins by explaining the fundamental workings of Large Language Models (LLMs). LLMs predict the next word in a sequence based on a given prompt, processing one word at a time. This process lacks a "back button," meaning LLMs traditionally don't have the ability to edit, reflect, or refine their output in multiple iterations like humans do when writing or planning.

To address this limitation, the concept of "compound LLMs" is introduced. A compound LLM involves stringing together multiple LLMs to perform different tasks in sequence, leading to higher quality results.

Example:

• LLM 1 (Draft): Writes a draft marketing plan based on a detailed prompt.
• LLM 2 (Critique): Critiques the draft marketing plan, identifying areas for improvement.
• LLM 3 (Update): Updates the original draft based on the critique, generating a revised plan.

This approach allows for reflection and refinement, mimicking the iterative process humans use when creating content. The video references a Google prompt engineering study where adding "please take a pause between each step of the answer" improved accuracy, highlighting the benefit of reflection in AI tasks.

Agentic AI: Beyond Compound LLMs

The video then transitions to the concept of "Agentic AI," building upon the foundation of compound LLMs. To understand Agentic AI, the video proposes three mental leaps:

1. LLMs as Agents: Instead of referring to LLMs, think of them as "agents" performing specific tasks. In the marketing plan example, each LLM (drafting, critiquing, updating) is an agent.
2. Agents as Tools: Not every agent needs to be an LLM. Some agents can be tools like Google Search, API calls, or calculators. For example, a "data request agent" could identify data needed to support a presentation, and a "Google Search agent" could retrieve that data.
3. Dynamic Workflows: The sequence of steps doesn't need to be predetermined. An "orchestrator agent" can dynamically adjust the workflow based on the task's needs and available tools.

Example:

Expanding on the marketing plan/presentation example:

• Data Request Agent: Identifies statistics needed to support the presentation.
• Google Search Agent: Searches for the required statistics.
• Draft Agent: Incorporates the statistics into a new draft.
• Critique Agent: Critiques the updated draft.
• Orchestrator Agent: Determines whether more data is needed or if the draft is ready for finalization, potentially looping through the data request, search, and draft stages multiple times.

The orchestrator agent has overall instructions, knowledge of available agents and tools, and a general workflow, but it can adapt the process based on the results of each step. This is analogous to the interactive "Bandersnatch" episode of Black Mirror, where the audience's choices influenced the narrative.

Conclusion: Embracing Agentic AI

The video concludes by emphasizing the potential of Agentic AI to create more flexible and intelligent AI systems. By combining LLMs with other tools and using orchestrator agents to manage dynamic workflows, organizations can develop AI solutions that are more adaptable, efficient, and effective. The speaker encourages viewers to explore Agentic AI and consider how it can be applied within their organizations.