A Primer on Agentic Systems

The world of artificial intelligence is at a pivotal crossroads. Large Language Models (LLMs) like GPT-4 have dazzled us with their ability to generate human-like text, assist in drafting emails, write code, and even compose poetry. But as remarkable as these models are, they primarily function as passive responders — waiting for input before offering output. Imagine, however, a new breed of AI that doesn’t just wait to be prompted but actively engages with its environment, makes decisions, and takes actions to achieve goals. Welcome to the era of Agents.

In this first installment of our blog series on agentic AI, we’ll explore the exciting transition from LLMs to agents and understand why this evolution is necessary.

Evolution of Gen AI Architectures: From LLMs to Multi-Agentic Systems (MAS)

To better understand the evolution from simple Large Language Models (LLMs) to sophisticated agentic systems, let’s explore this diagram depicting the progression of AI architectures in terms of complexity (of usage) and capability. The diagram visualizes how AI systems evolve in functionality as we move from basic LLMs to more advanced multi-agentic systems.

Let’s examine each of these systems in detail to understand how they work.

A Basic LLM System: Working with the World Knowledge

The diagram shows how a Basic LLM System works — a simple AI model that generates text based on user queries.

1. User Query: The process begins with the user providing a query — like a question or prompt.
2. World Knowledge: The LLM uses pre-trained world knowledge (data from books, articles, etc.) to understand and respond to the query. This knowledge is static, meaning it doesn’t update in real time.
3. Generated Text Sequences: The LLM produces text responses that are coherent and relevant based on its training. It’s a passive responder — it can only generate text, without taking any actions.

This basic architecture is the foundation for more advanced AI systems that can act, adapt, and interact beyond just responding with text.

These systems are positioned at the lower end of both complexity and capability, Basic LLM Systems represent the foundational AI models that generate text responses based on pre-trained, static world knowledge. They are passive responders — providing outputs when prompted, without any real-time interaction or action-taking abilities. This simplicity makes them useful for static content generation tasks but limits their ability to engage dynamically.

A RAG LLM System: Enhanced Knowledge for Better Responses

The RAG LLM System (Retrieval-Augmented Generation) represents an upgrade from basic LLMs. It’s designed to be smarter and more relevant by combining world knowledge with localized information to produce even more accurate responses. This simplified diagram shows how a RAG LLM system works.

The Retrieval-Augmented Generation (RAG) LLM System takes the capabilities of basic LLMs a step further by integrating localized, real-time knowledge through information retrieval. This allows the system to access up-to-date information from external sources, making the responses more contextually relevant and better informed.

Unlike basic LLMs, RAG systems can deliver enhanced responses by combining pre-trained world knowledge with specific, retrieved data — making the answers more accurate for specific situations.

However, despite these improvements, RAG LLMs are still passive responders. They do not autonomously take actions or make decisions, lacking the proactive features of more advanced AI systems.

The Agentic LLM System: From Generating Text to Taking Action

Imagine an AI system that doesn’t just answer your questions, but also takes action based on those answers. The Agentic LLM System is designed to do exactly that — it goes beyond simply generating text and moves into the realm of acting autonomously. Here is a simplistic depiction of these Agentic LLM systems work:

Let’s break down what’s happening in this diagram:

1. Knowledge Sources: The system taps into both world knowledge (general information) and localized knowledge (specific, current data). This means it can use both what it already knows and newly retrieved data to make decisions that are both informed and relevant.
2. Agentic Capabilities: The Agentic LLM goes beyond just generating information. It has the ability toinitiate actions. After processing the user’s query and combining all relevant information, it can determine the necessary course of action, carry out a sequence of steps to accomplish the task, and complete the intended goal through these actions. This allows the Agentic LLM to interact with its environment, going beyond being just a passive information provider.
3. Environmental Context and Feedback: The system doesn’t operate in isolation — it can understand and respond to environmental context. After taking an action, it also receives autonomous feedback — meaning it can understand how effective its actions were and adjust if necessary. This feedback loop helps the system learn and improve over time.

Let me explain with a simple example.

Imagine you’re using a smart assistant to control your home. You say, “Set the perfect mood for movie night.” A basic LLM would just give you suggestions, but the Agentic LLM System would:

• Understand the Context: Know what kind of lights, temperature, and sounds make for a great movie night.
• Take Action: Dim the lights, adjust the temperature, and turn on your sound system.
• Adapt: Based on your feedback, it could make adjustments to improve the experience next time.

Multi-Agentic System (MAS): When AI Agents Work Together

Imagine not just one AI helping you out, but multiple AIs working together, talking to each other, and solving problems as a team. This is what a Multi-Agentic System (MAS) is all about — a group of intelligent agents cooperating in a shared environment to achieve goals.

Let’s break down the diagram step by step:

• Multiple Agents and Their Interactions: In the Multi-Agentic System, you have multiple agents (e.g., “agent 1,” “agent 2,” etc.) all working together within the same environment. Each agent has its own role and strengths, just like how different members of a team have their own expertise. These agents are capable of agentic interplay, which means they constantly collaborate, negotiate, and even adapt their behaviors based on the situation. For example, Agent 1 might negotiate with Agent 2 on how to divide a complex task, and Agent 2 might adapt its behavior depending on Agent 1’s needs.
• Multi-Step Actions: What makes MAS powerful is that these agents can perform multi-step actions. In the diagram, you can see how they can initiate an action, execute the steps, and ultimately achieve an outcome. This is like a group of teammates deciding on a game plan, executing each step, and achieving victory together. By working in sync, these agents can handle much more complex tasks than a single agent could ever achieve alone.
• Emergent Behavior and Environmental Feedback: The real power of MAS lies in the emergent behaviors that arise when these agents collaborate. By working together and interacting, they can solve problems more efficiently and discover creative solutions — just like how brainstorming in a group often leads to better ideas than working alone. MAS also uses environmental context and autonomous feedback to adjust its actions in real time. For instance, if these agents are managing a smart home, they can adjust the temperature or lighting based on real-time changes like weather or occupancy.

Let us deep dive into a MAS scenario:

Consider a full-scale marketing campaign for a new product launch. A Multi-Agentic System (MAS) can orchestrate this by deploying specialized agents that work together seamlessly. Here’s how it works:

1. Content Strategy Agent: Agent 1 develops the content strategy by analyzing customer data, competitors, and trends. It determines the optimal content mix across social media, blog posts, and videos.
2. Copywriting Agent: Agent 2 is responsible for writing engaging content. Drawing from Agent 1’s strategy, it crafts blog articles, social media posts, and email newsletters that align with the brand’s voice and resonate with the target audience.
3. Design Agent: Agent 3 handles visual design, creating compelling graphics, infographics, and social media visuals that complement Agent 2’s copy. It maintains visual consistency with the brand image and marketing objectives.
4. Social Media Scheduling and Engagement Agent: Agent 4 optimizes content scheduling across platforms to maximize engagement. It monitors interactions, responds to comments, and collects user feedback.
5. Performance Analytics Agent: Agent 5 tracks real-time performance data, analyzing metrics such as engagement rates, clicks, conversions, and shares. When performance dips, it coordinates with other agents to refine the strategy.

How Multi-Agent Collaboration Works:

• The content strategy agent coordinates with the copywriting and design agents to maintain campaign coherence.
• Agentic interplay enables collaboration and negotiation. For example, when the performance analytics agent identifies low engagement, it recommends adjustments to the copywriting agent or design agent.
• Agents continuously adapt their behavior based on real-time feedback, ensuring the marketing campaign’s ongoing effectiveness.

Now that we have explored the LLM evolution, let us understand why do we need these agents.

Why Do We Need Agents?

One question that often arises is: If Large Language Models (LLMs) have become so powerful, you might wonder:

Why can’t one model handle everything we need?

Well, there are some key reasons why we need something more advanced than just a single LLM to perform complex actions. Let’s break it down.

1. LLMs Work Well with Prompts, But Not Always with Complex Actions: LLMs are great when you give them a prompt or question, like “What’s the capital of France?” But when you need them to complete multiple actions in a sequence, things get complicated. Imagine you ask the model to plan your trip, book tickets, order a cab, and send reminders — all at once. A single prompt like that could easily confuse the model, leading to mistakes and ineffective outputs.
2. Traditional Methods Fall Short for Multi-Step Tasks: To handle more difficult tasks, people have tried using techniques like chain-of-thought prompting, where the model thinks step-by-step. While this can work for task-based scenarios (like solving a math problem), it’s still not enough for real-life situations where many steps and decisions are involved.

To solve these challenges, we need agents — AI systems that can handle complexity and uncertainty. Here are some reasons why agents excel in action-based tasks:

1. Scalability: Agents can break big problems into smaller, manageable pieces. This makes it easier to solve complex tasks. Imagine organizing an event — an agent can split it into smaller tasks like finding a venue, sending invitations, and arranging catering, and then handle each task efficiently.
2. Robustness: Agents are designed to keep working well even if something goes wrong. Let’s say one part of a system fails (like a robot that temporarily loses its internet connection) — agents can adapt and continue operating without falling apart.
3. Flexibility: Unlike traditional systems, which need to be reprogrammed whenever things change, agents can adapt to new situations on their own. For example, if a delivery route becomes blocked, an agent can adjust and find an alternative path without waiting for someone to fix the program.
4. Efficiency: Since agents can operate autonomously (meaning they don’t need constant supervision), they can complete tasks much faster than human-operated systems. Imagine an agent managing your smart home — it turns off lights when no one’s around, adjusts the temperature, and makes decisions based on your habits — all without you having to ask.

Conclusion

As we’ve explored in this primer, the evolution from traditional LLMs to agentic systems represents a significant leap forward in AI capability. These agents, whether working independently or in multi-agent systems, offer unprecedented levels of autonomy, adaptability, and problem-solving abilities. They address the limitations of conventional LLMs by breaking down complex tasks and operating with greater flexibility and robustness.

In our next blog post, we’ll dive deeper into the specific characteristics that make these agents unique and explore various architectural patterns that enable their sophisticated behaviors.

Stay tuned to learn more about the building blocks that power these intelligent systems.