Agentic AI: Introduction and Journey to new operating paradigm

Since late 2023, the landscape of Applied AI has undergone a dramatic transformation. The emergence of AI agents marks one of the most profound paradigm shifts—expanding the early success of Large Language Models beyond text generation and conversational applications to a far broader canvas of real-world applications.

Enough and more has been written on this topic across millions of articles and blog posts, with no dearth of material on the internet and social media — ranging from enthusiastic optimism to skeptical caution. The emotional spectrum spans awe, excitement, and curiosity on one end, to fear, concern, and even resistance on the other, as both strong arguments for and against the rise of AI Agents continue to fuel the ongoing debate.

One often wishes there were fewer options—just enough to avoid the cognitive overload of having to choose.

About this series

As mentioned in my announcement post for this series, my goal is to move beyond the hype and focus on documenting and exploring both the technical architecture and practical applications of AI agents—specifically tailored to enterprise needs. The intent is to equip enterprise leaders, architects, and implementers with the clarity and depth needed to understand the potential and pitfalls, digest key system-level details, and confidently embark on this multi-year transformation journey—overcoming skepticism and caution along the way. Some of the topics in my mind as I take the journey of writing this series are as below:

1. Core design primitives — planning, memory, tool-use, control, and learning, model optionality, governence needs, agent learning: data/knowledge loops, multi-turn trajectories with actions. multi-agent systems and more.

2. Next level design patterns in each of those and options and trade offs.

3. Architectures for scalable, safe, and productive agent systems. Role of Platforms in accelarating, building, orchestrating and deploying agents at scale in enterprise (I would like to show you glimpes of Xceed AI agents Platform)

4. Real-world lessons from building agentic applications in enterprise use-cases.

5. Open challenges that still need answers, hardwork, skills

6. How does the future direction look? Given we are still on an exponential curve with rapid advancements on application and system side of AI Agents

This Post

Since this is the first post in the series, some repetition is inevitable. This is to ensure completeness and provide context for future readers who may want to revisit this series later. Feel free to skip sections that seem familiar.

Transformation Journey from LLMs to Agentic AI

Like every journey, there is always a starting point and we embark on the path with a destination in mind.

The original strength of LLMs lay in their remarkable ability to memorize and compress the world’s knowledge. What surprised us most was how human-like they felt—we could hold conversations, ask complex questions, summarize articles, generate blogs like this one, write code, and unlock countless other applications.

At the heart of it all was a powerful new capability: Generative intelligence

2023–2025+ — The Rise of Agentic AI and the new scaling Laws

As the LLMs, felt human like in their conversational ability, often sounding like experts and very often dumb, but still compelling enough to be part of our daily work such as writing, seeking knowledge, etc. The next natural question emerged: can they also do work which involves actions ? Can they take on real-world tasks?

This requires more than just language—it demands the ability to reason, plan, break down complex goals into manageable steps, and interact with the external world through tools and APIs. Moving from generative models to truly useful agents means enabling them not just to respond, but to act.

Sanjeev Mohan and I were among the early voices to highlight this transition in our post on 2024 2024 Data and AI Trends and we reaffirmed it in our 2025 Enterprise Data and AI Trends. As I write this, I'm confident that this journey is only beginning and will continue to unfold in the years ahead.

This is a multi-year transformation—shifting today’s IT systems from rule-based, GUI-centric software agents to conversational agents powered by Large Language Models.

Between 2023 and 2025, we’ve seen the rapid rise of activities Agentic AI—AI systems that go beyond generation to reason, plan, take actions, and interact meaningfully with the world. We entered the Agentic era — where AI Models(LLM) don't just answer but act: ✅ Autonomous task execution ✅ Goal-oriented planning ✅ Learning from feedback & evolving over time

While early Large Language Models impressed with their generality, the new frontier is task specialization: We need agents that can learn and adapt to specific tasks, environments, and user goals.

While one appreciates the drivers behind this transformation, It's extremely important to note that this change is acting as a forcing function across many different aspects. To limit the length of this post, I want you to zoom in and critically review the following important tenants at the top of the transformation needs.

1. How do we learn? Transition from Model Learning to Agent Learning

2. How do we build? Transition from Rule based Agents to LLM based agents

3. How do we operate and govern? Transition from IT GRC to Human worker like GRC

Learning --- Moving from Model Learning to Agent Learning

While the previous learning paradigm including base transformer architecture, Supervised Fine Tuning/DPO that helped align model to base human preference and capabilities and formed the bedrock of the foundational learning and reasoning. Agentization needed an additional newer learning paradigm which enabled models to learn from feedback, optimize long-horizon multi-turn needs for complex, goal-directed tasks.

Clearly, we needed to move from static models which were the outcome of current batch learning mode to dynamic learners, helping them improve through trial, correction, and interaction, not just memorize generic bookish concepts, but also learn to do real world tasks in a real task environment, and overcome the learning needs of often more nuanced, complex, messy tasks.

This journey from building learning systems which seem human like knowledgeable to building systems that learn and act in real world environments is not only reshaping AI research but also how we build, train, deploy and monitor these new software systems (or learning systems), which can go to work alongside us humans, is what is shaping the agentic era post 2023 and it is likely to be a multi-year transformational journey.

In a sense, its no longer about how much the model knows (model size), but in how capabilities scale with task complexity, tool use, memory depth, and autonomy.

"The future of AI is no longer just about building bigger models—it's about creating smarter, more adaptable learning agents capable of picking up a wide range of human tasks in real-world environments."

Building - From Fixed Logic, IT Systems to Adaptive Intelligence

Our Information Systems Landscape today largely includes rule based systems, in some cases, narrowly learning decision systems (built on classical Machine Learning) wrapped in an existing software system.

LLM based agents represent a functional shift in how we approach building next gen software system, automation systems and decision making system.

Traditional rule-based software operates through explicit, predefined conditional logic -- if-then statements that cover anticipated scenarios but often struggles with edge cases and requires constant human update as requirements evolve. There's often an elaborate workflow and user-interface layer which wraps these applications. These systems are built in mind for integrating human workflows. These workflows include decision workflow, information workflows, also automation workflows, to take away predictable very well defined human work.

However, largely these systems acted as dumb tools that were built to help humans reason and act. In contrast, LLM-based agents enable and interpret natural language instructions, reason through complex scenarios, and adapt to novel situations without explicit programming for every contingency. They have integrated tools to enable them take actions. This paradigm shift enables more flexible, context-aware systems that can handle ambiguous inputs, engage in multi-step reasoning, and maintain conversational interactions with users.

The goal of new agent systems is not just aid humans to reason and act, but could they offload workflows that require to handle uncertainties, and take more action on behalf of humans, act as an assistant to human worker, with human worker taking the supervisory role and instructing these agents to carry out real action.

In a sense we are also taking away the predictability of rule based systems for enhanced adaptability and emergent problem solving capabilities in a less predictable, semi defined workflow and environment.

This transition is not binary. It has a spectrum of needs depending on the organisation, use-case, need for control/oversight. Organizations making this transition must choose and carefully balance the benefits of increased flexibility with the need for appropriate oversight and validation mechanisms.

Operating & Governing - IT Systems GRC to Human Worker Like Governance and Control

Earst while software systems typically followed a waterfall approach where developers would define comprehensive logic, test thoroughly, deploy once, and maintain through periodic updates when new rules were needed. Ofcourse, even there, the need for speed moved us to more agile shorter and faster build to deploy cycle.

Classical Machine Learning and the original LLM learning paradigm operated on a batch learning paradigm where models are trained on historical datasets, validated through cross-validation and holdout testing, then deployed as fixed artifacts that serve predictions until the next scheduled retraining cycle—typically triggered by performance degradation or new data availability.

LLM-based agents, however, introduce a more complex hybrid approach that combines pre-trained foundation models with continuous adaptation through techniques like fine-tuning, retrieval-augmented generation, and in-context learning.

Unlike classical ML's periodic batch retraining, LLM agents can adapt through prompt engineering, few-shot learning, and dynamic retrieval without traditional model retraining, enabling real-time behavioral modifications.

The deployment complexity also escalates significantly:

1. Rule-based systems require simple application hosting, classical ML needs model serving infrastructure with version control, while LLM agents demand sophisticated orchestration platforms that can handle multi-step reasoning, external tool integration, memory management, and real-time prompt optimization.

2. How do you onboard an agent to work, may look similar to how you onboard an employee, train and deploy it but with supervisory/governance controls in place.

3. We are used to human workers and the governance processes around it have evolved over multiple centuries.

4. The evolution of agents as digital workers reflects a shift from static logic to statistical learning, and now to adaptive reasoning. This progression challenges us to rethink the governance and control frameworks of current IT systems—shifting from software-centric oversight to models that more closely resemble employee governance, with autonomy, accountability, and oversight by design.

Conclusion: Navigating the AI Agent Transformation

The spectrum of human response to AI agents—from profound excitement and curiosity to legitimate fear and resistance—reflects the magnitude of this technological shift. Both perspectives are understandable, as each side draws from real data points and experiences that validate their concerns or enthusiasm. However, the greatest obstacle to successfully navigating this transformation often lies not in the technology itself, but in our own cognitive frameworks—the judgments, experiences, and belief systems that shape how we interpret and respond to exponential change.

Transformational technologies consistently challenge our ability to accurately predict their trajectory and impact. We tend to underestimate the rate of improvement and overestimate our capacity to maintain the status quo. For decision makers and IT executives, this AI agent revolution represents more than a technical upgrade—it's a fundamental reimagining of how intelligent systems operate and evolve. Rather than viewing this complexity as a barrier, consider it an invitation to exploration.

The organizations that will thrive in the coming years are those that approach this transformation with the mindset of an explorer: curious rather than certain, adaptive rather than rigid, and willing to experiment rather than wait for perfect clarity. The journey from rule-based systems to intelligent agents is not just a technological migration—it's an organizational learning expedition that requires both courage and strategic thinking. The question isn't whether this transformation will happen, but whether you'll be leading it or reacting to it.