In the world of artificial intelligence, the term "agent" has become a buzzword. Many people believe AI agents are magical solutions that can solve all problems just like Aladdin's Genie. But the truth is, AI agents come in different forms, each with its own strengths and limitations.

Let me explain with a story.

It’s a busy Monday morning at a luxury car company Build Your Car (BYC). Managing Director, Mr. Thompson, walks into the office and hands his Personal Assistant (PA) Sarah a list of tasks. 

1.      Sarah: The Reliable Personal Assistant

First, he asks Sarah to answer five specific questions about production delays and draft an email to a dealer. Sarah doesn’t hesitate. She cross-references schedules, pulls data from reports, and crafts a polished email, all in 20 minutes.

What’s happening here?

Sarah is reactive - she waits for clear instructions and then acts swiftly within her rulebook. No creativity is needed - just precision. 

This is your classic AI Assistant: think chatbots or email organizers. They’re great at tasks like:

o   Answering specific questions (“What’s Q3’s sales?”)

o   Drafting emails, scheduling meetings, or sorting data

o   Following rules exactly as programmed

2.    Cook: The Travel Agent

Later, Mr. Thompson asks Sarah to book a 5-day business trip to Turkey. She forwards the request to Mr Cook, company’s travel agent, who follows a strict workflow: check flight availability, compare hotel rates, calculate costs, and confirm bookings. 

What’s happening here?

The travel agent uses predefined workflows. Think of it as a recipe: follow steps A → B → C, and you’ll get a predictable result.

This is a Workflow System: powered by AI tools like LLMs. They’re reliable for tasks like:

o   Processing insurance claims

o   Generating reports from templates

o   Automating customer service tickets

3.     David: The Strategic Sales Head

Enter David, the new Sales Head. Mr. Thompson throws him a curveball: “Increase our revenue.” No step-by-step guide. No predefined rules. David digs into market trends, analyses competitors, brainstorms promotions, and proposes a bold plan: partner with ride-sharing companies to target fleet buyers.

What’s happening here?

David is proactive - he doesn’t wait for instructions. Instead, he identifies opportunities, weighs risks, and makes decisions. 

This is an AI Agent: the free-thinking problem-solver. These systems:

o   Set their own goals (e.g., “Boost revenue by 15%”)

o   Use external data (market trends, customer behaviours)

o   Experiment with strategies and learn from outcomes

Now that we've seen how AI Assistants, Workflow Systems, and AI Agents work in a real-world scenario, let's summarise the key takeaways: 

§  AI Assistants = Sarah. They follow orders.

§  Workflow Systems = Cook. They follow recipes.

§  AI Agents = David. They think, adapt, and strategise.

The Power of System Design in AI

Large Language Models (LLMs) on their own are limited by the data they've been trained on. This limitation impacts their knowledge of the world and the types of tasks they can solve. However, LLMs are still useful for various tasks, like summarising documents, drafting emails, and reports.

LLMs have limitations. To unlock their full potential, we must integrate them into structured systems and existing workflows. But what does this mean? The term "system" implies multiple components. Systems are modular by design. We define our approach based on desired outcomes and select the right components to solve complex problems. Additionally, we can define the system approach based on what we desire our system to achieve and select the necessary components, tools, and workflows to solve complex problems.

In the first use case, Sarah needs access to production delay data and then pass it back to the LLM to draft an accurate email. 

Similarly, in the second use case, Cook needs access to the destination and days from the calendar and then use a few APIs from reservation sites. We can then pass the result to the LLM to create an itinerary along with a cost estimate. 

Both of these examples illustrate compound AI systems. Notably, we don't need to fine-tune the model or employ anything overly complex. A well-defined programmatic control flow with an established workflow can solve these cases.

Introducing AI Agent

Now, let's discuss where agents fit in. One way to manage the logic of a compound AI system is to use an LLM for reasoning and decision-making while integrating automation for execution. This approach is made possible by the tremendous improvements in the reasoning capabilities of large language models like OpenAI-o1 or DeepSeek-R1.

LLM can be fed complex problems, and you can prompt them to break these problems down and come up with a plan to tackle them. Another way to think about it is to consider two ends of the spectrum. On one end, you're telling your system to think quickly, follow established protocols, and avoid deviations. On the other end, you're designing your system to adopt a more thoughtful and intentional approach.

This involves creating a plan, attacking each part of the plan, identifying potential roadblocks, and adjusting the plan as needed. In the face of complex questions, there's often a temptation to provide immediate answers. However, this approach often leads to incorrect solutions. Instead, breaking down the problem, identifying areas where external help is needed, and taking the time to solve it carefully can lead to higher chances of success.

This thoughtful and intentional approach is what we mean by an agentic approach, where a large language model is put in charge of the control logic.

Components of AI Agents

Let's break down the components of an agent, a key part of the Agentica architecture:

§  Reasoning: The ability to analyse problems, develop plans, and evaluate each step in the process. This involves putting the model at the core of problem-solving.

§  Action: The ability to execute solutions through external programs, known as tools. These may include searching the web, accessing data sources, interacting with APIs, executing functions and more.

§  Memory: The ability to access and store information, including inner logs and conversation history. This enables the model to retrieve and build upon previous knowledge, providing a more personalised experience.

Make a Wise Choice

When building LLM-powered applications, we recommend starting with the simplest solution and adding complexity only as needed. This might mean not building agentic systems at all, as they may introduce unnecessary latency and cost. Agentic systems often trade latency and cost for better task performance, and you should consider when this trade-off makes sense.

When more complexity is warranted, workflows offer predictability and consistency for well-defined tasks. In contrast, agents are the better option when flexibility and model-driven decision-making are needed at scale. 

Think back to BYC’s transformation journey. Initially, their systems functioned like Assistants - relying on humans to provide context and make decisions. But as they moved toward automation, they needed something more dynamic, capable of reasoning and acting autonomously. This shift required well-structured agentic design patterns to ensure reliability, efficiency, and adaptability. Let’s explore these foundational patterns.

Key Agentic Design Patterns

There are many ways to configure an AI agent. The agentic design pattern introduces common solution patterns that make LLMs more autonomous and enhance AI model performance through strategic thinking, tool integration, and collaboration, ultimately creating more sophisticated and reliable applications. According to Dr. Andrew Ng, a globally recognised leader in AI, four key agentic design patterns shape the performance and reliability of LLMs in complex tasks. We'll explore these four patterns:

a)            Reflection-Based Reasoning

The Reflection Pattern focuses on improving AI's ability to evaluate and refine its own outputs. This self-reflection loop is not limited to a single iteration. The AI can repeat the reflection process as many times as necessary to achieve a refined, polished result. By using this approach, the model's accuracy and reliability can be enhanced through self-guided corrections.

Consider BYC's marketing team creating a brochure for one of their upcoming luxury cars. The team collaborates with an LLM to generate the first draft. Initially, the team reviews the draft, provides feedback, and incorporates the changes. This process is repeated until the brochure meets the desired standards.

Now, imagine this process with the Reflection Pattern. Instead of the team reviewing and providing feedback on the draft, the LLM itself evaluates the content, identifies areas for improvement, and revises the draft. The LLM essentially plays the dual role of creator and reviewer, iteratively refining its own output until it meets the desired standards.

This autonomous self-reflection enables the LLM to learn from its mistakes, adapt, and produce more accurate and reliable content, ultimately creating a high-quality brochure that showcases BYC's luxury car in the best possible light.

b)           Tool-Use Reasoning

The Tool Use Pattern substantially expands an LLM's capability by allowing it to interact with external tools and resources to enhance its problem-solving abilities. Instead of relying solely on internal computations or knowledge, an AI following this pattern can access databases, search the web, or even execute complex functions via programming languages like Python. This pattern is powerful because it allows AI systems to tackle more complex, real-world tasks where internal knowledge alone isn't sufficient, expanding their utility into real-world applications.

Suppose BYC's Sales team is trying to generate a report on the latest sales trends in collaboration with LLM. Without the Tool Use Pattern, the LLM would be limited to its internal knowledge and training data, potentially resulting in outdated or incomplete insights. However, with this pattern, the LLM can access external resources like databases, web APIs, or other data sources to gather the latest data, analyse it, and generate a more accurate and up-to-date report. This enables the LLM to provide more informative and actionable insights, demonstrating the value of the Tool Use Pattern in real-world applications.

c)           Planning-Based Reasoning

The Planning Pattern enables an LLM to break down large, complicated tasks into smaller, more manageable components. Planning enables the LLM to react to requests and strategically structure the steps needed to achieve a goal. Instead of tackling a problem in an ad-hoc manner, an LLM using the Planning Pattern will create a roadmap of subtasks, determining the most efficient path to completion. This method yields higher-quality results and consistency, especially in complex problem-solving.

One notable implementation of this pattern is ReAct (Reason + Act), which combines reasoning and action to achieve a goal. ReAct enables LLMs to reason about a problem, create a plan, and execute the plan to achieve the desired outcome.

Consider this time BYC’s Marketing Team is asking for an analysis of competitor pricing, market trends, and customer preferences to formulate targeted promotions. Using the Planning Pattern, the agent breaks down the task into smaller subtasks:

o   Analysing competitor pricing for luxury cars in BYC's target market

o   Identifying market trends, such as seasonal fluctuations in demand for specific models

o   Determining customer preferences, including favourite features and colours

o   Formulating personalised promotions, such as exclusive discounts, free upgrades, or VIP test drives

o   Scheduling and deploying the promotions through targeted marketing channels, including email, social media, and in-dealership events

By using the Planning Pattern, the Market Analysis Agent can create a tailored sales strategy that resonates with BYC's valued customers, driving sales and revenue growth.

d)           Multi-Agent Reasoning

The Multi-Agent Pattern facilitates collaborative problem-solving by leveraging multiple specialised LLM instances. This pattern builds upon the concept of delegation, similar to human project management. By assigning different agents distinct roles, each agent can handle various subtasks independently while communicating and collaborating with others to achieve a common goal.  

There are several types of multi-agent patterns. Collaborative Agents, where multiple agents work together on different task parts. Supervised Agents, where a central agent manages and verifies others' activities. Hierarchical Agents, where higher-level agents oversee lower-level agents in a structured system.

If we recall, Sarah and Cook’s tasks fit well within structured workflows, ensuring efficiency and predictability. However, David’s challenge is entirely different - it demands adaptability, reasoning, and strategic decision-making. A predefined workflow won’t work here. Instead, Multi-Agent systems is more appropriate comprising multiple specialised LLM instances:

§  Sales Trend Analysis Agent: analyses current sales trends, identifying strengths, weaknesses, and areas for improvement

§  Market Analysis Agent: analyses market trends, competitor pricing, and customer behaviours

§  Inventory Analysis Agent: examines existing stock/inventory, identifying opportunities to optimise levels and reduce waste

§  Customer Insights Agent: analyses the customer base, including demographics, vehicle age, and purchasing trends

§  Strategic Sales Agent: weighs inputs from other agents, developing a comprehensive sales strategy to drive revenue growth

In this example, each agent plays a distinct role, collaborating to achieve the common goal of increasing revenue. The Multi-Agent Pattern enables the BYC to leverage each agent's strengths, leading to a more effective and innovative solution.

Agentic Framework

Agentic Frameworks provide structured approaches to designing and developing autonomous agents or multi-state systems. These frameworks simplify the development of complex agent systems by handling low-level tasks such as calling LLMs, defining and parsing tools, and chaining calls together. While they create abstraction layers that can obscure underlying prompts and responses, they make development easier.

Frameworks like LangGraph, CrewAI and AutoGen can help you get started quickly. However, as you move to production, don't hesitate to reduce abstraction layers and build with basic components to ensure transparency, control and scalability. It's also crucial to thoroughly understand the underlying code to ensure effective debugging.

Conclusion

As we've journeyed through the world of AI agents, it's clear that their potential extends far beyond mere automation. By harnessing the power of agentic design patterns, we can create intelligent systems that learn, adapt, and innovate. The future of AI lies not in standalone models, but in integrated systems that combine reasoning, automation, and modularity. As we continue to push the boundaries of AI, it's essential to refine our approach, embracing the complexity and nuance of agentic design to unlock truly intelligent solutions.