I am so excited about the subject of AI agents and agentic workflows that I’m making this topic a four-part series. Today we begin by defining what AI agents and agentic workflows are, and why they’re game-changing.
Admittedly, when I bring up AI agents in casual conversation I’m in for a few blanks stares. For some this topic won’t have obvious curb appeal. Hear me out.
Over the last two years, regardless of our role in the business, many of us received a crash course in AI technology and its potential to transform the business landscape. While some of us have already begun to see meaningful changes in how we work, most have yet to experience the seismic shifts in operational and organizational design that have been predicted.
This might explain why the fever-pitched conversations around AI have subsided somewhat. Regardless, those following the conversation about AI developments in places like The World Economic Forum know that possibly the most consequential advancements are currently underway with the advancing capabilities of AI agents and the development of agentic workflows.
To the lay observer, advances in AI agents might appear as merely the expected incremental progress on now-familiar tools. In reality, if the buzz is to be believed, the rise of Large Language Models (LLMs) and Generative AI may be contextualized most notably as critical links in a chain leading to this moment—the emergence of fully realized AI agents and agentic workflows which promise a new way of working, a new era in business, and create the potential to fundamentally shift economies.
Am I getting ahead of myself? Maybe. It’s easy to do when talking about AI. So, let’s take this from the beginning.
What are AI Agents?
Simply put, an AI agent is a form of artificial intelligence designed to perform specific tasks autonomously. It can gather information from its environment, process information, make decisions, and take actions to achieve predetermined goals.
Chatbots are a form of AI agent you’re probably familiar with. They can vary greatly in sophistication. Earlier forms, which you’ll still find in use, meet only the very basic criteria for an agent by our current standards. These chatbots primarily use scripts to respond to queries and have limited capabilities.
Alternatively, the most advanced chatbots today are much more capable; answering far more complex questions, maintaining context within and across interactions/conversations, learning from interactions with users, adapting their responses based on new information, integrating with other systems to provide targeted answers and taking action toward solving issues.
So what made this leap forward in AI chatbot technology possible? According to reporting from Cade Metz and Karen Weise of The New York Times, the answer is Minecraft, the wildly popular video game.
When OpenAI released GPT-4, and Nvidia got access to its underlying technology they saw the potential for it to be capable of do more than just text creation. To explore these possibilities, Nvidia’s team of researchers began developing the first LLM-powered AI agent named Voyager. The goal was to create an autonomous Minecraft player with abilities on par with a human player.
The Voyager site page on MineDojo explains that Minecraft was chosen as an ideal testing ground because, unlike other AI games, it doesn’t impose a predefined end goal or storyline but provides a context where endless possibilities can play out. This environment allowed them to test whether an autonomous AI agent could propose suitable tasks, refine skills, commit mastered skills to memory, and seek new tasks independently.
Metz and Weise concluded, “The project was an early sign that the world’s leading artificial intelligence researchers are transforming chatbots into a new kind of autonomous system called an AI agent. These agents can do more than chat. They can use software apps, websites, and other online tools, including spreadsheets, online calendars, travel sites, and more.”
In business, AI agents of increasing sophistication are used to execute tasks across a wide range of functions. In marketing, for example, AI agents are used in all aspects of the funnel—from analyzing large volumes of customer data to segment audiences based on demographics, to optimizing ad campaigns by adjusting bids, and even identifying suitable influencers for collaborations. Increasingly, AI is used to generate creative content.
Here we can see how AI agents are linked along a funnel in the fashion of a daisy chain (my term); a conceptual prototype leading us to the evolution of agentic workflows.
What are Agentic Workflows?
Agentic workflows are an approach to organizing multiple AI agents into a cohesive, iterative workflow directed and managed by an organizing entity.
If we use the analogy of a football team to represent an agentic workflow, the organizing entity would be the team coach. The coach chooses the players (AI agents), assigns the positions (order in the workflow), and creates the plays (the workflow itself).
Agentic workflows are defined in part by having this central organizing AI; ensuring different AI agents work together harmoniously and leveraging their collective strengths. The organizing entity can coordinate the tasks, manage data flow, and integrate feedback loops, resulting in a more efficient and effective overall process.
Feedback loops are another defining feature. The key innovation of agentic workflows is an embedded iterative process. Through advanced prompt engineering techniques, agentic workflows embed iterative processes that allow for continuous improvement and refinement of outputs. This contrasts with traditional "one-shot" prompting, where AI generates a single response without further refinement.
Speaking at Sequoia Capital's AI 2024 summit, Andrew Ng (thought leader, co-founder of Coursera and Google Brain Project) spoke on the transformative potential of agentic workflows in AI development.
“Agentic workflows involve AI models engaging in iterative processes, delivering remarkably better results compared to non-agentic workflows.”
Ng defined four key design patterns essential to agentic workflows:
Reflection: The AI reviews and improves its own output, much like a human would revise their work.
Tool Use: AI leverages external tools and data sources to enhance its capabilities.
Planning: AI agents autonomously plan their actions, improving strategic task completion.
Multi-Agent Collaboration: Multiple AI agents collaborate, each performing different roles, to tackle complex tasks more effectively.
The moment that stood out the most to me in Mr. Ng’s presentation was the revelation that within workflows agentic components can become more than the sum of their parts.
According to Ng, “GPT-3.5, when used within an agentic workflow, outperformed GPT-4 in certain coding tasks by leveraging iterative improvements and collaborative problem-solving.”
This means that even as we await new releases of some of the most powerful AI tools put forth by OpenAI, Meta, and Google et. al., the potential of conceived agentic workflows to drive significant advancements in AI is widely and currently available; offering a new paradigm that combines iterative refinement, strategic planning, and collaborative problem-solving to achieve even better results. This puts entrepreneurs and creative engineers in control of supercharging and personalizing AI tools and capabilities in bespoke and meaningful ways.
The excitement around this moment in AI evolution is well-founded, driven by the tremendous implications for the future of business and creativity. For this reason the AI-Curious Newsletter will be taking a deeper dive into AI agents and agentic workflows in the coming months.
Stay tuned! In Part 2 of our series, I’ll be talking to Near Futurist, Neil Redding, of Redding Futures. Neil is an author, keynote speaker, and innovation architect. His talk at SXSW earlier this year fundamentally changed the way I think about spatial computing and the concept of “being in a place.” I can’t wait to get his thoughts on the potential impact of next-generation AI agents and the near future of agentic workflows.
Awesome!