In a previous post, we have already explained in broad and rather abstract terms the capabilities promised by agentic AI systems. Now we will (in a slightly lighter tone) look at the capabilities that perhaps one of the most popular chatbots and assistants, ChatGPT, has, and specifically examine the shortcomings that separate it from the world of agentic AI.

ChatGPT is one of today’s best-known Artificial Intelligence tools that can convincingly mimic human language use. Many people are inclined to think that “ChatGPT can do almost anything,” but of course this claim should be taken with a grain of salt. ChatGPT belongs to the category of Large Language Models (LLMs), which, although highly advanced in text understanding and generation, still differ significantly from so-called agentic (or autonomous) AI. The essence of agent-based AI is that

• the ability to set goals autonomously,
• develop a strategy to achieve them, and then
• take the necessary actions without the need for constant human guidance or intervention.

ChatGPT does not have such capabilities. In practice, it acts more as a kind of “language assistant”: we ask it questions or give it tasks, and it responds to them. However, if we don’t speak to it, it doesn’t initiate its own topics, write its own emails or try to come up with new ideas – it’s a completely passive and reactive system. You could say that it lacks proactivity and initiative. This passivity is the result of a conscious design decision, which is beneficial from both a security and a user perspective. The developers wanted to create a tool capable of processing huge amounts of linguistic information and generating natural-sounding responses, but which does not perform autonomous, unsupervised operations.

That’s why ChatGPT does not automatically search for new data, run in the background, or set up new sub-topics for itself to investigate. This is a limitation, but it also ensures that the system remains under the user’s full control and never acts at his or her own discretion. Agent-based AI, on the other hand, differs in its ability to operate autonomously and achieve goals autonomously, but this also raises several new opportunities and potential problems.

A truly autonomous system can receive a remote goal – for example, “map potential business partners in Europe” – and then break down the request into fine-grained subtasks. This can be done, for example, by creating its own “prompts” or code calls for each subtask and starting to execute them autonomously. All the while, it does not require constant human intervention. If it gets stuck at a step or encounters a conflict, it can reroute the execution path, look for more information and adjust its strategy accordingly.

Imagine what it would be like to leave an Artificial Intelligence system on your computer for two or three days. During this time, the system would scan a variety of external data sources, send emails, create spreadsheets and draw conclusions – all autonomously and without human control. This kind of self-directed, proactive task solving is the essence of agentic systems: they not only respond to user queries, but can take the initiative, adapt and move towards the goal themselves.

An example of a rudimentary agentic AI is the AutoGPT Agent, which is already available as a pilot project in the GPT store of OpenAI. It is a system that can perform some of the described functionality independently. It breaks down the instructions received from the user into parts, creates its own prompts, executes them one by one, and finally combines the results into a coherent response. This approach provides an opportunity to go beyond the traditional “one question one answer” level of operation of ChatGPT.

It is important to note, however, that AutoGPT is still immature in many respects. While it shows a promising direction, it also faces practical and security constraints that cannot be addressed before it becomes a true agent-based AI system. In addition, in many cases, even the execution of subtasks is not automatic, although they only need to be approved, but user intervention cannot be avoided completely. Nevertheless, such experiments clearly demonstrate that the industry is no longer just developing passive conversational partners but is getting closer to creating autonomous digital agents. This transformation has huge potential, but also a serious responsibility to ensure that autonomous AIs remain safe and ethical.

The short answer to the question “Why isn’t ChatGPT agentive?” is that the system doesn’t yet have the basic capabilities that are essential for autonomous operation. For example, it doesn’t have a true extended memory system that can retain context over the long term and link it to current actions. It also doesn’t have the mechanism that would allow it to initiate its own actions on external devices or platforms. Each of its responses is based solely on instructions given by the user, whether it’s writing code, analyzing data, or browsing the web. If these limitations exist, ChatGPT (and LLMs like it) remain extremely smart but essentially passive assistants.

However, this does not mean that ChatGPT is not an extremely useful or innovative tool. Generative language models are already revolutionizing content creation, education, and everyday problem solving, even if in their current form they are more of a productivity tool. Many experts are working to make these models more “agentic” with additional plugins, API integrations, or specialized modules. However, the system still needs a lot of development to achieve autonomous operation.

ChatGPT is a significant milestone in the development of Artificial Intelligence, offering solutions in a variety of areas with its exceptional language capabilities. However, it is worth remembering that this technology may only be the starting point for the future in which agent-based AI will help us in our daily lives. If we are aware of exactly what it knows and what it cannot do, we can more realistically exploit its potential. We can also have a clearer picture of the direction in which Artificial Intelligence needs to develop to create truly autonomous systems.

István ÜVEGES, PhD is a Computational Linguist researcher and developer at MONTANA Knowledge Management Ltd. and a researcher at the HUN-REN Centre for Social Sciences. His main interests include the social impacts of Artificial Intelligence (Machine Learning), the nature of Legal Language (legalese), the Plain Language Movement, and sentiment- and emotion analysis.