A relatively new term is being widely used in media and social media these days – ‘Agentic AI’. I say relatively because the term has been around for more than a year, but not in everyday or consumer space. How rapidly is the field of AI evolving? Rapid enough for terminologies to shift within months. Earlier software technology evolved at the pace of several years per major update. Last weekend a couple of students and a group of professionals reached out to ask whether ‘Agentic AI’ is a new type of AI, which is why I am addressing the term this week.
So what is Agentic AI? It is an updated term to reflect newer capabilities in existing Generative AI products. Till a few months ago, Generative AI products such as ChatGPT, Gemini, Claude etc were called AI Assistants and prior to that they were called LLMs or chatbots (or both). When Generative AI products were first introduced, they were what I call ‘respondents’, tools that respond in text-based answers to questions or requests for information. At this stage they were called ‘LLMs’ or ‘chatbots’ – LLMs to reflect that the products were based on Large Language Model architecture and chatbot (Chat+robot) to reflect the concept that they were automated chat programs. As Generative AI as a field evolved, the products expanded their capabilities from single or simple information retrieval tasks to more complex work that involved multiple tasks in different sequences. For example, summarising information or structuring content or providing a table of contents in response to a user’s request. At that stage, the same products were called (positioned as) ‘Assistants’ since they were now capable of providing a wider range of outputs. Now – as in today – the same product has added even more capabilities to independently perform job roles involving some amount of autonomous decision making along with executing several complex processes each requiring one or more tasks. For example, today ChatGPT, Gemini, Claude and the like can independently carry out research from online sources, evaluate gathered information, analyse and write a report based on a methodology. To reflect these emerging capabilities, companies have coined the term ‘Agentic’ to signify the ability to perform autonomously or independently. Thus, a class of Generative AI tools has evolved from performing limited robotic functions to performing processes under orders and with guidance to independently performing specific job roles. And this has happened in the user-facing product within a span of 2 years. A lazy (but quick) analogy would be: At 11, I was a writer who expressed himself in personal diaries, by 16 I was able to write formal essays and long-form articles for school and inter-school competitions (which meant I was takin into consideration external evaluation parameters); By 21, I was proficient at writing researched reports for decision makers at work (so added the capability of weighing business impact of research and make predictions); By 27 I was a feature writer and columnist for the city edition of a national newspaper (added competency to write, analyse and predict impact after weighing social political, cultural and geographical considerations). Same me but different capabilities within the same skill-set (writing).
Before a term is introduced to the customer ecosystem, it usually appears in internal or technical literature such as documentation and white papers. The same can also be traced for the evolution of the term Agentic AI.
Following is a timeline of terms and related capabilities as they evolved (incidentally, the information was generated with the assistance of ChatGPT).
| Phase | Period | Product Term | Backend Term | Capabilities |
| LLM/Chatbot | 2020-Q3 2023 | LLM, Chatbot, Text Generator | Langauge Model, Completion AI | Single-turn Q&A, static text generation, no memory, no tool use, passive response to prompts. |
| Assistant | Q4 2023 – Q4 2024 | Assistant, Copilot | Tool-using model, planner, context manager | Provide assistance on specific tasks, integrates with tools, limited context awareness, reactive, limited memory. |
| Agentic AI | Q1 2025 – Present | AI Agent, Agentic AI, Autonomous Assistant | Agent loop, agentic behaviour, multi-step planner | Autonomously plans, execute tasks across tools, adapts via memory, reasons over time horizons, completes goals involving nutlike task with little or no supervision, collaborative digital agent. |
As a user of multiple Generative AI tools, I can verify this information based on my experience (have documented research on the same). When I look at conversations from early 2024, I can see the limited functionality and remember thinking that many of the behaviours of the platforms were like 5-year-olds. By November 2024, my impression had changed because it was in October 2024 that most of the platforms released the updated models which reflected ‘reasoning’ abilities. In fact initially I used to find funny because you could see messages like “Thinking for 2 seconds” or “Reasoning for 3 seconds”. My instinctive reaction was: “Who says ‘I am thinking for 5 seconds’?”.
Many of the platforms still have that visible reasoning but the messages have been upgraded to “Assessing”, “Reading resources”, “Analysing”and so on. These are like the machine’s notes to self, telling itself what it is doing and I personally think this is lazy on the part of User Interface designers and product programmers to leave it like that. I find ChatGPT more intuitive and less pretentious on this particular aspect. Ask a question and except for ‘searching” it just responds.
But the capabilities of Generative AI platforms are impressive now. And this is where it can become interesting: if one thinks, looks at the evolution of the internet, of software programming, of AI and then analyses – we can make predictions of what all might come next. But that’s for a different day. For this blog, the objective was to examine Agentic AI and clarify that it is not a new type or classification of AI or Generative AI but a term that https://srisabaripackersandmovers.com/contact-us/ symbolises the maturing capabilities of a particular class of Generative AI: the multimodal LLMs. At this point, Audio, Image and Video based Generative AI are not Agentic AI because they are mainly generative and lack the ability to plan or make decisions. For Audio, Image and Video based Generative AI to become Agentic they would have to be integrated with an LLM. I could be wrong but that’s where the technology is at this point.