Agentic AI - My take

Introduction

In recent months, Agentic AI has emerged as a focal point in the technology sector, captivating both industry leaders and innovators. This paradigm shift moves beyond traditional AI's reactive nature, introducing systems capable of autonomous decision-making and proactive task execution.

Several major companies are at the forefront of developing and implementing Agentic AI solutions like OpenAI ,Microsoft, NVIDIA Corporation, GitLab Inc. to name a few.

So let's talk about -

What is Agentic AI :

The world is witnessing a rapid evolution of artificial intelligence (AI), with a key player emerging: AI Agents. These sophisticated software programs are not mere tools; they are intelligent entities capable of autonomous action, learning, and adaptation. From automating mundane tasks to revolutionizing entire industries, AI Agents are poised to transform how we live and work.

At their core, AI Agents are designed to perceive their environment, make decisions, and take actions to achieve specific goals. Unlike traditional software, which relies on predefined rules, AI Agents leverage machine learning algorithms to learn from data, adapt to new situations, and continuously improve their performance.

How is it related to LLMs & Generative AI :

Large Language Models (LLMs) are a specific type of generative AI that specialize in understanding and generating human language. The LLMs are subset of Deep Learning.

You want to know more about LLMs. Read my article https://www.dhirubhai.net/pulse/large-language-models-vs-small-raja-saurabh-tiwari-kiljf/

Generative AI is a broad category encompassing AI systems that can create various forms of content, including text, images, music, and code.

LLMs are a subset of generative AI, focusing on text-based outputs, while generative AI encompasses a wider range of creative content generation capabilities.

Examples:

• Text-to-image models: DALL-E 2, Midjourney, Stable Diffusion ?
• Language models: GPT-3, Bard, LaMDA
• Music generators: AIVA, Amper Music ?

Agentic AI

Agentic AI goes beyond creation. It's about autonomous action and decision-making. These AI systems can operate independently, set goals, and take actions to achieve them. Imagine a self-driving car – it perceives its environment, makes decisions, and navigates accordingly.

Examples:

• Self-driving cars
• Robotics in manufacturing
• AI-powered trading algorithms
• Virtual assistants

Generative AI is like a composer writing a symphony. Agentic AI is like a conductor leading an orchestra, orchestrating the performance and ensuring it achieves its desired effect.        

Okay, so now you have understood the basics of the different terms. Let’s go little deeper into Agentic AI world.

The idea of Agents is not new, “Shakey” was the first mobile robot with the ability to perceive and reason about its surroundings. https://www.sri.com/hoi/shakey-the-robot/

Earlier it was more of static rule based systems which evolved to goal-driven intelligent agents which happened in 80's.

Okay so what are agents ?

“Agents” can be defined in several ways. Some define agents as fully autonomous systems or components that operate independently over extended periods, using various tools to accomplish complex tasks. Others use the term to describe more prescriptive? implementations that follow predefined workflows. So we need to understands the distinction between workflows and agents.

Workflows are systems where LLM and tools are orchestrated through predefined code paths.

Agents, on the other hand are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.

So there are basically 3 stages I would say

Automation, AI workflows and then Agentic Workflows

Automation is simple automation that we work day to day like automatically generating email triggers when finding an error in the log

AI workflows are a level up, where multiple such automation tasks are combined with ML and AI capabilities and orchestrated by some code.

Some examples could be,

Workflow:

Example: Writing a marketing email campaign.

1. User Input : A user enters a prompt: "Generate a blog post about the future of AI in healthcare."
2. Text Generation: A Generative AI model (e.g., GPT-4, Claude) generates a static blog post based on the given prompt.
3. Response: he generated text is presented to the user as-is.
4. No Iterative Refinement: The model does not evaluate its own response or attempt to improve it based on user feedback.
5. Manual User Intervention Required: If the user wants modifications (e.g., "Make it more technical"), they must manually edit the prompt and regenerate the content.

Agentic AI Workflow (Focus: Autonomous Action)

Example: Self-driving car navigation.

Workflow:

1. Sensor Data Acquisition: The car's sensors (cameras, radar, lidar) continuously gather real-time data about the surrounding environment (traffic, pedestrians, road conditions).
2. Perception & Interpretation: AI algorithms process sensor data to understand the environment.
3. Decision-Making: The AI agent analyzes the perceived environment and makes real-time decisions.
4. Action Execution: The car executes the decisions by controlling the steering wheel, accelerator, and brakes.
5. Continuous Learning : The AI agent continuously learns from driving experiences, adapting its behavior to improve safety and efficiency.

Key Differences

AI Workflow: Primarily focuses on content creation, often requiring human oversight and refinement.

Agentic AI Workflow: Emphasizes autonomous decision-making and action within a dynamic environment.

The below diagram illustrates how the agentic workflow works. It depicts some key concepts of the Agentic workflow.

But how do I start, are there any rules, any patterns, templates which I need to follow to create agents? Let's take a look at the design patterns.

Agentic Design Patterns :

The agentic design patterns at its core talk about how the agents collaborate, use environment and make autonomous decisions. These are very key in writing scalable solutions.

• Multi Agent Collaboration : The goal is divided into multiple tasks, which means multiple agents are created based on their expertise. These agents collaborate to generate the results/achieve the goals. Agents communicate via message passing or shared state (blackboard model).

Planner Agent: Generates high-level goals.

Execution Agent: Performs tasks based on plan instructions.

Critic Agent: Validates actions and suggests improvements.

Memory Agent: Retrieves past knowledge for context-aware responses.

• Reflection : The Reflection pattern enables AI agents to analyze their past actions, assess performance, and refine future decision-making autonomously. This is crucial for continual learning and self-improvement without explicit retraining. This is like an retro that we do towards end of each sprint.

Key Features:

Self-Evaluation: Agents assess their performance using metrics or goals.

Feedback Loops: They adjust their strategies based on the evaluation.

Scenario: A recommendation agent reflects on user feedback to improve its suggestions.

The agent logs its past experiences and outcomes. It uses feedback loops (e.g., reinforcement learning reward signals, heuristic evaluations). It performs meta-cognition (thinking about its own thinking) to optimize strategies.

• Tools Used : AI agents often integrate external tools and APIs to extend their abilities beyond language generation, such as executing code, retrieving knowledge, or interacting with real-world systems.

Retrieval & Memory

Execution & Automation

• Planning : The Planning pattern allows AI agents to strategically sequence actions to achieve a goal, optimizing efficiency and handling uncertainty.

It is evident that one of the first step of planning would be 'Goal decomposition'. Followed by path selection. These paths should be adjusted dynamically based on the each iteration's output.

These Agentic AI design patterns enable intelligent systems to become self-improving, collaborative, and goal-driven, making them more capable of autonomous problem-solving in real-world applications. By leveraging reflection, multi-agent coordination, planning, and external tools, developers can create AI solutions that continuously learn, adapt, and enhance decision-making capabilities.

Conclusion :

So in my view the agentic AI is another layer of automation. This takes it to the next level by having autonomous agents doing their job, challenging the response, improving the results and then finally giving their best results to the client.

Everyone is talking about creating AI agents in the market, there are lot of tools available for creating agents, but creating an agent which solves the real world problem is hard today.

Always remember to have simpler solution to solve the business problem than focusing on using a technology which might end up having complex solution. You should think if a given business problem can be solved by simple workflows. Then prefer that.

Having said that, while we are at the very early stages of the Agent realm, evaluation and tracking are essential for an accurate and responsible response.

In my next article, let's have a walkthrough of creating AI agents from scratch using python.

#naturallanguageprocessingnlphashtag#nlphashtag#machinelearninghashtag#artificialitellegencehashtag#datasciencehashtag#textanalytics

Thanks,

Raja Saurabh Tiwari