I bet we all ask the same question when we hear about AI: Is it smart enough to understand our needs, and if so, how?
The answer is an absolute big YES. If you have tried using the DeepSeek assistant or ChatGPT (o3 model) recently, you would have seen the model's fun and complicated reasoning process, just like a human. This is an example:
We call that “Agentic Reasoning”.
Agentic Reasoning
Agentic reasoning is the ability of modern AI systems to break down complex tasks into manageable steps and then execute those steps autonomously. In simple terms, agentic reasoning allows your digital assistant to answer a question but also plan and act on your behalf.
Isn’t this how we humans do?
You are planning a family dinner.
You send the invitations to everyone.
You then create a grocery list.
You search for the recipe for the dishes you want to cook for the family.
That’s how AI reasoning and chain of thoughts work, too.
These are all tasks that, when done manually, can take up a surprising amount of time and mental energy. Thanks to agentic reasoning, AI tools today can break down a big, complicated task into smaller, more actionable pieces (sometimes even better than what a human can do!).
For example, you ask AI to provide you with a grocery list. Instead of just responding with “Here is your grocery list,” the AI agent can analyze your dietary preferences, check what ingredients you already have at home, suggest recipes, and generate a detailed list of what you need to buy.
This kind of multi-step planning wasn’t possible with older AI systems.
From Simple Responses to Autonomous Planning
At the beginning of AI hype, the tools were primarily rule-based systems that could only perform pre-programmed tasks. You would ask a question, and the AI will give you a canned response. No real “thinking” was involved; it was more like looking up an answer in a vast database. That’s why earlier systems were great for simple, one-off queries but couldn’t help with more complex, everyday tasks.
With the advent of advanced models like OpenAI’s o1 series and others, AI systems can engage in chain-of-thought reasoning. This means they can generate intermediate steps that help them arrive at a solution, much like we do when planning our day. With this approach, AI tools can now assess a problem, plan a series of actions, and execute them step by step - all with minimal human intervention.
What’s Enabling Agentic Reasoning Today?
Two significant factors have made agentic reasoning possible:
Advancements in Machine Learning and Natural Language Processing (NLP):
Today’s AI systems are powered by sophisticated models that learn from vast amounts of data. These models better understand and generate human language and maintain context over longer interactions. They “think” through problems by mapping out a chain of reasoning before delivering a final answer.
This is a massive leap from earlier generations' more straightforward, template-based responses.
Increased Computational Power and Improved Algorithms:
The hardware supporting AI has evolved dramatically. Modern GPUs, specialized AI chips, and cloud-based infrastructures now provide the immense processing power needed for these complex tasks. This computational muscle, combined with refined algorithms, allows AI systems to perform real-time planning and decision-making.
Companies like Google, Microsoft, and NVIDIA are investing billions in these technologies, which means the cost and speed of executing these tasks have improved enormously.
Three Main Stages of Agentic Reasoning
The technological foundation of agentic reasoning is built on three clear stages:
Plan Stage:
Use natural language processing and large language models to understand your query
Identify key details and provide the proper context.
Refine your query with domain-specific knowledge to ensure precision.
Retrieve Stage:
The refined query is matched with the right enterprise data.
The system fetches accurate and relevant information using role-based access controls, semantic search (keyword and vector-based), and knowledge graphs.
If it doesn’t find a satisfactory result, it returns to the planning stage to refine the query further.
Generate Stage:
Synthesizes all the retrieved data into a clear, human-readable response using advanced generative models.
This response is personalized based on your context, past interactions, and specific needs, with citations included to ensure transparency.
More and more AI companies will utilize these reasoning processes to develop powerful agents with knowledge of specific business contexts. Such products could be code debugging agents, cloud deployment agents, unit test writing agents, legal document drafting agents, and marketing materials producing agents, to name a few. As a software engineer, I am looking forward to those tools that can make every engineer a 10x engineer. I shall explore these tools to see how they can help our daily work.
That’s it for today’s post. Cheers.
In my next post, I will introduce a tool called Replit Agent, which enables anyone to enter a prompt to create a software project from scratch. It's both intimidating and exciting.
Subscribe to “A Curious Semi-Expert” newsletter to receive my weekly learnings directly in your inbox. I will keep you company as we adapt to these exciting technologies together.