After the novelty of generative artificial intelligence (AI) wore off, many raised an important question — Yes it is cool, but how can it make an impact in the real world? It was a valid question. While AI chatbots can be seen as a one-stop-shop for quickly looking up information, having an impromptu conversation, making it write essays, generate images or videos, their role is largely limited to a system where a human user will have to constantly command it to get an output and oversee the result.
Even if its capabilities cannot be dismissed, and it did make a significant impact in improving workers’ productivity in certain areas, it lacked one critical element that stopped it from becoming a faithful assistant that could handle and truly automate tasks — decision-making. Generative AI today can help with certain aspects of a person’s work, but it cannot execute a task.
For instance, you can ask it to write an email to a client letting them know about an unexpected delay, but it cannot send that message or deal with the angry reply they send. Similarly, you can use Gemini or ChatGPT to ask for “the best smartphone for shooting videos”, and it can recommend the latest iPhone 16 Pro Max or the Samsung Galaxy S24 Ultra. But it will not be able to scour the web to find you the best deal and make a purchase.
Realising this gap, tech companies working on large language models (LLMs) began using the word AI agent. Researchers believe AI agents can take a knowledge-based AI system and transform it into an action-taking system that can perform end-to-end tasks without human intervention.
The term gained prominence during the second half of 2024, and currently, it is being treated as a panacea for all work-related problems. And while there is some truth to it, is it really a transformative technology of that potential? The answer might be a bit complex, but we will do our best to break it down and highlight all the different aspects that you should know about. Let us dive into it.
What is an AI Agent?
Since this technology is still in its nascent stage, there is no unified definition of what exactly constitutes an AI agent. IBM defines it as a system “that is capable of autonomously performing tasks on behalf of a user” by designing a workflow and using tools. Similarly, Google, which announced its first AI agent dubbed Project Mariner last year, calls it a system that acts like an assistant to humans and helps them complete tasks.
A more comprehensive definition is given by Amazon, which describes it as “a software programme that can interact with its environment, collect data, and use the data to perform self-determined tasks to meet predetermined goals. Humans set goals, but an AI agent independently chooses the best actions it needs to perform to achieve those goals.”
Put simply, an AI agent can be understood as an AI system that can take action instead of just telling the user about the action.
Breaking Down the AI Agent
A typical AI agent will have a large language model (LLM) as its brain. But it will also include other elements that enable it to use that intelligence in actions. Most commonly, these extra parts are different sensors, mechanical parts, encoders, or integration into other software.
The sensors enable an AI agent to collect data across different formats. These can be visual, sound, temperature, or electronic signals. Mechanical parts are typically used for embodied AI or robots which need to execute real-world actions such as lifting an object or moving from one spot to another. Encoders are used to convert different types of signals into information that can be processed by LLMs. Finally, software integration enables the ability to execute tasks.
It is also important to highlight another crucial difference between AI models and AI agents at this point. AI models contain a pre-training database which forms the basis of their knowledge. Anything that is not part of the database will not generate an output. A good example of this was the early version of ChatGPT which was not connected to the Internet and had a knowledge cut-off date. If it was prompted to answer a question about current affairs, it would not be able to answer that.
On the contrary, AI agents, when integrated with relevant systems, can independently collect new data to solve problems that would not be possible based on their existing database. For instance, Google’s Project Mariner can interact with the browser to find the best deal on a smartwatch.
Another aspect of AI agents is the capability to handle complex tasks. AI agents are capable of advanced reasoning and as such can break down a complex task into multiple easier tasks and then complete them one after another. This contextual understanding of the problem and the ability to know how to break it down is a fundamental function of AI agents.
A good example of this is Gemini’s recently added Deep Research tool. Users can ask it to explain a technical or niche topic. The AI would then create a multi-step research plan, break down the topic into smaller parts, find relevant research papers and articles on the topic, execute the plan, conduct research, and analyse the gathered data to create a detailed report.
Applications of AI Agents
AI firms have been touting AI agents as a tool that can be used across industries and in different scenarios. It can be used as a voice assistant for devices that can perform device-specific tasks (such as taking a picture or playing music). It can be added to an app or software and carry out tasks within that (purchasing a product via a browser-based agent). It can also be added to enterprise systems and it can detect fraud or find ways to optimise different processes.
Apart from this, AI agents are also said to perform transformative tasks in certain industries. In healthcare, it can be used for diagnosis, treatment recommendation, and drug discovery. In the automotive sector, it can be used to create self-driving cars. AI agents are also said to be able to pilot drones in disaster areas to gather and analyse data and offer actionable insights for rescue operations.
It also has applications in manufacturing industries via AI-powered robots, in the gaming industry as a game developer or as a non-playing character (NPC) inside games, and in the education sector to create personalised study plans and to grade test papers in a human-like fashion.
However, it is important to note that while tech companies are marketing AI agents as a catch-all for all kinds of end-to-end intelligent automation, the current technology limits its use case to largely specific task-based roles instead of a general-purpose tool.
AI Agents in 2025
With that being said, it is important to ground our expectations and understand what we can realistically expect from AI agents in the current year. It is unlikely that AI agents will enter the workforce in any of the critical sectors such as manufacturing, automobiles, healthcare, or education.
However, this year should mark the entry of AI agents in consumer electronics, mobile and desktop applications, as well as websites and platforms. Google’s Project Mariner, for example, could be integrated with Google Chrome and assist users in making purchases and finding files from the web by the end of this year.
OpenAI is also rumoured to launch its AI agent this year which could further enhance ChatGPT’s capabilities and allow it to perform certain actions on a user’s device and the Internet. Anthropic’s Computer Use tool is also expected to make a global release and assist users in their day-to-day tasks on the device.
Eventually, we should also see a shift where AI agents can mimic keystrokes, mouse movements and clicks, and do much more on devices. For example, by the end of the year, more agentic tools such as the coding agent Devin could be writing code end-to-end, testing them, finding and fixing glitches, and deploying them without human intervention. But, it would be highly optimistic to include this in the 2025 itinerary.
On the enterprise side, AI agents could take a larger role in completing some organisational tasks such as monitoring a large volume of data, preparing analytical reports, and offering recommendations and course corrections. It could also be used in some cybersecurity roles. Notably, Meta has stated that it already uses AI to ensure guidelines are being followed. YouTube also uses AI to monitor copyright violations.
However, we do not expect AI agents to enter any of the critical work functions this year because the technology is largely untested and its reliability will be questionable. Businesses, specifically public enterprises or those backed by large investors, are generally risk-aversive and are unlikely to provide access to sensitive data.
The Problems With AI Agents
With AI being the current trend in the tech space and the potential to disrupt a large number of industries, it is understandable why there is so much excitement about AI agents. However, beyond the rose-tinted glasses exist several issues with AI agents that need to be addressed before the technology can witness large-scale adoption. On the other hand, if it goes unchecked, the technology can pose several risks.
One of the main issues with AI agents is bias and discrimination which comes from their training data and can lead to discriminatory outcomes. This also highlights another issue of transparency in AI agents. With complex algorithms and architecture, most AI agents are complicated and opaque systems where it is difficult to understand how and why they make certain decisions.
There are security and privacy issues as well. From a security perspective, AI agents can be vulnerable to adversarial attacks, where malicious actors manipulate input data to deceive the system. Additionally, since AI agents need to be connected with multiple systems and collect a large amount of data to carry out tasks, they also pose privacy risks.
With so many challenges, AI firms will have a tough job ahead to convince enterprises and individuals of the upside of the technology while reassuring them of the downside. Regardless, it cannot be denied that AI agents will constitute a large part of AI announcements in 2025.
