🦸🏻#3: Illustrating Agentic Vocabulary: Real and Potential Examples

The previous episode of our agentic series, where we offered a go-to vocabulary, has become our most popular premium subscription post. Nothing surprising – terminology around AI agents is vague and often misleading. People use whatever terms they want, especially in the business world. “Build me a bot!” – a business inquiry – might actually mean a multi-framework agent that connects different databases, provides communication, and performs analysis. While the vocabulary we created is great (though not perfect, of course), we received feedback: “Show some concrete examples to illustrate this.” Thank you, dear reader; you are absolutely right. We couldn’t provide examples along with the vocabulary due to the newsletter’s size restrictions, but we are doing it now.

However, we have to warn you – it will be brief, because to become aware of AI agents, one needs just to look around. We all have been using AI agents for many years now. We can barely imagine our lives without them.

But why do we talk so much more about AI agents now? Yup, blame – or rather thank – generative AI. LLMs enabled a giant leap towards more autonomy and perceived intelligence in AI agents. Now, with multimodal models improving exponentially, AI agents are rising to a new level.

Still, that’s not the only thing that LLMs improved. LLMs have also revolutionized how we interact with AI. Beyond zero-shot prompting, LLMs enable agentic workflows, allowing AI to think*, research, revise, and improve iteratively. (*Please consider all antropomorphizing verbs to be in””). This evolution transforms agents into more adaptable and capable entities.

The pinnacle of these workflows is multi-agent systems, where multiple agents collaborate, breaking down complex tasks into manageable subtasks. This structure enhances control, testing, and maintenance while enabling sophisticated, scalable workflows like parallel or sequential task execution. Although challenges arise, such as the stochastic behavior of LLM agents increasing error risks, multi-agent architectures are a natural evolution. They align with modular development patterns and provide reusable, adaptable solutions, paving the way for the next generation of autonomous systems.

In today’s episode:

As a reminder, all AI agents fall under the broad category of autonomous agents, defined by their capacity to operate independently. They perceive their environment, make decisions, and take actions to achieve specific goals. These agents differ based on features such as learning capability, physical form, specialization, or task complexity. Intelligent agents learn and adapt, while rational agents aim for optimal outcomes given their knowledge. Task-oriented agents, bots, and simple agents focus on specific tasks, with varying levels of adaptability. Smart agents operate in dynamic environments, adjusting based on feedback. Embodied agents have a physical form/body, while digital agents operate solely in virtual spaces. AI assistants, copilots, and AI personas interact with humans, offering varying degrees of specialization and personalization. This categorization shows that autonomy and task complexity define an agent’s sophistication, not just its form.

It's important to remember that these categories are not mutually exclusive. Modern AI agents rarely fit neatly into a single box. The terms might overlap significantly, exhibiting characteristics from multiple categories, leading to hybrid forms with diverse capabilities.

Now, let’s take a few very routine things we use every day. They turned out to be agents:

Let’s admit, chatbots are still most of the time frustrating and dumb, but thinking about a smart toilet as your AI assistant is fun. Using the vocabulary we provided you with, you will be able now to distinguish what is what.

Now, to some more advanced stuff.

Agentic workflows are gaining traction as a more effective way to leverage AI by enabling systems to operate in structured, iterative processes, surpassing the limitations of traditional, single-prompt AI methods. These workflows are designed to manage complex tasks by breaking them into smaller, manageable components, allowing AI agents to interact dynamically, learn from their actions, and adjust strategies based on real-time feedback.

Andrew Ng has been one of the most prominent voices in promoting and explaining this concept. One striking example from Andrew Ng's research shows that even GPT-3.5, when used in an agentic framework, outperformed zero-shot GPT-4 by 10-40% on certain tasks. This illustrates the power of the agentic approach: even older models gain a competitive edge when structured iteratively with proper planning and tool usage.

Agentic workflows offer different autonomy levels, ranging from basic rule-based automation to fully autonomous systems capable of managing highly complex, real-time scenarios. This scalability allows organizations to choose the level of autonomy that best fits their operational needs and risk tolerance​.

These workflows are used across industries, including logistics, eCommerce, and smart manufacturing. In coding, multiple AI agents work in tandem to write, review, and test code. In robotics, autonomous systems like drones follow agentic workflows to navigate environments, plan routes, and adjust based on sensory feedback​.

By integrating these components, agentic workflows enable AI systems to become more adaptive, autonomous, and capable of solving increasingly sophisticated problems, highlighting the evolution from simple automated processes to truly intelligent systems capable of collaborative problem-solving.

A few such advanced agentic systems, we mentioned last time. Let’s explore this triumvirate of terms that sound like they've been plucked from a sci-fi novel but are indeed rooted in our present and near-future reality: Multi-Framework Agents – Multi-tool Polyglots of AI; Self-Replicating Agents – The Digital Rabbits; Polymorphic Agents – The Shapeshifters. There is a lot of terms overlap that will help us grasp the interlinked essence of AI agents, how to talk about them, and how to build them.

These versatile agents operate across various platforms and formats, adapting seamlessly to different environments. They excel at integrating diverse data sources and operating across multiple technological ecosystems.

Multi-framework agents, like AutoGen, are designed to work across different platforms and environments, adapting their functionality and often collaborating with other agents to solve complex problems. AutoGen's multi-agent approach ensures more reliable and comprehensive responses compared to single-agent systems. For instance, in automated questionnaire answering, AutoGen employs specialized agents for answering questions, checking answers, verifying links, and managing the overall process.

Imagine a futuristic travel booking agent that takes the concept of multi-framework agents to new heights. This AI system would:

This advanced agent would create a seamless, personalized travel experience by harmonizing data from numerous sources and interacting with various systems.

Self-Replicating Agents possess the remarkable ability to create copies of themselves autonomously, a feature crucial for scalability and system resilience.

In the realm of cloud computing, Kubernetes Autoscaler and Amazon Web Services (AWS) EC2 instances demonstrate the power of self-replication. These systems:

Futuristic Nanobot Swarms (future)

Looking beyond software, imagine a future where self-replicating nanobots revolutionize medicine:

This futuristic application of self-replicating agents could transform healthcare, offering targeted, adaptive treatments at the cellular level.

Polymorphic agents adapt their behaviors based on context. Built using polymorphism principles from object-oriented programming (OOP), they share a common interface but specialize dynamically, promoting flexibility and scalability. This allows them to extend functionalities without changing core code, ideal for complex systems needing adaptability. Examples include agents for research, data collection, and fact-checking, all customizing methods while sharing a base class.

In smart factories, polymorphic robots demonstrate impressive flexibility. These robots can:

Envision a city-wide AI system that adapts its functionality to manage various urban challenges:

This polymorphic urban AI would seamlessly shift its focus and capabilities to address the ever-changing needs of a modern city, showcasing the potential of adaptive AI systems in complex, dynamic environments.

The advent of generative AI has supercharged everything. LLMs have pushed the boundaries of what we thought possible in natural language processing and generation, becoming the foundation upon which more specialized and capable agents are built.

Moreover, the rise of multimodal models marks another leap forward. By integrating understanding across text, images, audio, and video, these models bring us closer to AI agents that interact with the world in ways that mirror human perception and communication more closely.

While the terminology may sometimes seem as complex as the technology it describes, the underlying trend is clear: we're moving towards AI agents and workflows that are more versatile, adaptive, and integrated into our daily lives than ever before. It’s crucial to understand, however, that both simple and advanced terms don’t exist in isolation. They are part of a rich tapestry that includes more familiar concepts like AI assistants, copilots, and embodied agents. The field of artificial intelligence is less a strict hierarchy and more a Venn diagram with significant overlaps. But if someone asks you to build a bot, most likely he doesn’t know what he is talking about ;)