Are True AI Agents Even Possible?

Artificial Intelligence agents represent a significant and rapidly evolving branch of AI that is looking to disrupt industries through automation, improved decision-making, and operational efficiencies. However, the concept of an AI agent is often misunderstood. This paper aims to provide a precise definition of AI agents, while discussing their current limitations, including always having a human in the loop.

?

Definition of AI Agents

An AI agent is defined as an entity that perceives its environment and processes information to achieve a goal. AI agents expand these concepts using machine learning, generative models, and large-scale language models in order to perform more complex tasks and to interpret unstructured data.

Practically, AI agents automate activities previously reliant on human cognition. However, the promise of a fully automatic AI agent is still overstated. When I heard Sam Altman speak at Harvard in May 2024, he said that every major decision must be anchored in human responsibility and ethical guidance. It made me think more about how, no matter how smart AI gets, there’s always going to be a need for human oversight to keep things in check.

An AI agent depends on human intervention at many stages, from initial training to rule design and exception handling. The agent is fundamentally based on statistics and tendencies and is never based on real understanding. This human-in-the-loop is not a passing check; it will become a permanent feature, ensuring that human judgment and human values are never displaced.

Take for example the use of an AI Agent for tele-sales and business development. From the point of view of the salesperson who uses the AI Agent, this agent may appear, from his side, to be functioning independently; the agent’s scripted system, its response to its customers, etc. may all appear to be performed independently. But when we have the customer in mind, the perception of AI Agent appears completely false. The customer is constantly interacting with the agent, giving his or her opinions or inputs. The conversation provides the agent with information. The final behavior of the agent depends a lot on these responses, and it is impossible for it to operate in a way without such interaction.

?

Limitations and the Need for Human Oversight

In my opinion, true AI Agents are not possible, at least not soon. The idea of true AI Agents is driven largely by marketing rather than grounded in technological possibilities. Artificial intelligence, no matter how powerful it becomes, still lacks common sense and context awareness. It can excel in pattern recognition and probabilistic logic, but it depends on humans to set goals, comprehend end results, and solve all complications that do not yield to a simple programmatic algorithm.

In addition to the sales example, the same dependence is evident in other areas. In the financial markets, AI trading agents may place trades on the stock exchange independently, but asset managers and compliance officers provide the rules, supervise performance and ensure that the regulations are followed, and if any divergence appears, the agent is immediately closed down.

This means that advances in artificial intelligence are mostly focused on optimizing this collaboration rather than the pursuit of full autonomy. What is called independence in one aspect is based on deep dependence on the information from the other. Another aspect of the limitations is whose guardrails the various AI agents are to run in. Who sets them? It could be the developers, the researchers, but it is often the customers, both enterprises and end-users, who define the final requirements for each agent. This different layer of human involvement shows that AI agents do not ever arrive at their objectives independently; instead, they will continue to remain tied down to human judgments.

For example, AI agents are and will most likely always be constrained by human ethics, i.e. refusing harmful or inappropriate requests. An AI agent in theory should provide whatever output best matches its data. The fact that human oversight is required to prevent misuse, rather than just to enhance functionality, further illustrates the system’s dependency. Even as AI systems become more agent like in their capabilities, the importance of human oversight only grows. This growing need for alignment and ethics suggests that AI agents will never be fully built to be autonomous and independent, but instead shows the refinement of how AI and humans can co-create solutions that are dependable together.

?

What is Getting Developed in the Industry?

Despite these limitations, the concept of an AI agent is not all-or-nothing. Rather, it exists on a continuum, and this spectrum is where innovation and startup differentiation occur. Some young companies prefer the simpler models, which only classify data or find information; others aim further up, nearer to the agent by integrating the decision, the context, the capacity to take action based on human guidance. In this spectrum between limited, passive AI and the goal of a “true” AI agent, startups have the chance to distinguish themselves.

The startup landscape of AI agents is maturing rapidly. Over 200 deals were completed in the trailing 12 months, with a median deal size of $3 million, an increase of 50% year-over-year. Although total capital invested declined, the new deals that are happening shows that the market is shifting from general AI Agent companies to particular industry-specific companies.

The “AI Agent Startup Landscape” illustrates this trend, highlighting major players like OpenAI, Cohere, and Adept that have attracted substantial funding. Within this environment, startups that navigate the agent continuum effectively are poised to become the winners, whether they pursue it vertically or horizontally.

Tau Ventures is an early stage AI focused venture fund. Interested in learning more, reach out.