Finding AI’s Blockbuster App
Since the public splash around the release of ChatGPT, AI technology has made headlines and captivated the attention of the general public, IT specialists, and investors.
This has been especially true as the capacity of generative AI has been growing exponentially. In addition, a new wave of competition has accelerated progress in the field, with Chinese AI like DeepSeek challenging the cost structure and pricing of US-made AIs.
Still, the AI industry is somewhat on shaky ground, as the hundreds of billions in AI infrastructure are yet to generate the levels of revenues required to justify the investment.
Previous tech revolutions have been built on the back of profitable applications of the technology into the “real” economy, like office work (Windows & Office), entertainment (video game and streaming), ads (Google), communications (smartphones), or trade (online payment & e-commerce).
So far, AI has yet to revolutionize how most people work or live their daily lives. But this is likely about to change with the emergence of specialized AI with an exploding range of performances and abilities: AI agents.
What Are AI Agents?
The core idea of AI agents is to create AIs that can operate independently in a given environment. This gives it very different practical roles than the ones of generative AIs like LLMs (Large Language Models) or image generators, which are mostly reacting to human-created prompts.
In that context, “environment” can mean both specific situations in the real world, like a car on the road and AI for self-driving function, or a fully virtual place, like a specific software or digital interface.
Because the AI agent is acting autonomously in its given role, it does not require the constant intervention of humans via prompting. So, it can also take action by itself, without needing confirmation or supervision.
In practice, most AI agents will also have built-in conditions & rules where they are going to ask for feedback from a human supervisor.
According to Google, AI agents’ key features are :
- Reasoning: they can analyze data, identify patterns, and make informed decisions based on evidence and context.
- Acting: The ability to take action or perform tasks, this can include physical actions or digital actions.
- Observing: Gathering information about the environment or situation to understand their context and make informed decisions.
- Planning: Developing a plan to achieve goals, with identification of the necessary steps, potential actions, and the best course of action.
- Collaborating: Working effectively with others, whether humans or other AI agents.
- Self-refining: AI agents can learn from experience, adjust their behavior based on feedback, and continuously enhance their performance and capabilities over time.
Source: DevRevAI
Is It Really New?
This set of features puts AI agents one step above previous iterations of AI tools, like AI assistants and bots, with more proactive abilities, autonomy, and the ability to handle more complex tasks.
Source: Google
AI agents with physical “bodies” can interact directly with the real world, while digital AI agents are likely going to be specialized in certain virtual work environments.
In both cases, giving the AI enough agency and possible actions to be useful but not too much to avoid unexpected damages from errors is crucial.
Overall, it is likely that the parallel growth in the quality of AI decisions and the increasing familiarity with them will let people and authorities give more latitude to AI decision-making. This, however, opens interesting legal and ethical questions regarding the responsibility of AI actions (see below a discussion on that topic).
The Potential of AI Agents
The ability of AI agents to work independently is what could make them AI killer apps. Modern life is full of not-so-complex repetitive tasks that are at the same time too complex to be given to a simpler form of automation.
This is, for example, why Tesla had to take a step back in the automation and robotization of its assembly line in 2018. Robots could do a great job, but even the slightest disturbance or unexpected change in requirement would bring the whole assembly line to a halt.
“We had this crazy, complex network of conveyor belts … And it was not working, so we got rid of that whole thing. Yes, excessive automation at Tesla was a mistake. To be precise, my mistake. Humans are underrated.”
Elon Musk
However, modern AIs are not just a very elaborate set of rigid rules trying to anticipate everything in advance. Instead, they are able, to some extent, to adapt and evolve to new conditions if given enough relevant data during their training.
So this could make AI especially relevant for highly repetitive tasks, from walking a customer through a troubleshooting algorithm to driving trucks on a highway.
Contrary to humans, such AI could work 24/7 and not require salary, health insurance, etc.
There are many ways to classify the level we have reached in AI abilities. Overall, they tend to compare the ability of AI to the general human population, with the newest AI agents reaching maybe soon the skills of 50-90% of the population in specific, narrow-domain tasks, usually considered as a mid-point in AI progress, and just the beginning for AGI.
Source: Cobus Greyling
How AI Agents Work?
AI agents are made of a few key “components” interacting with each other:
- Sensors: for physical AI agents, this can include cameras, microphones, LIDAR, radio antennas, etc. For digital AI agents, this can be a search function, a tool to read files, extract data from a specific software or database, etc.
- Actuators: This is how the AI can perform its job. When physical, it can be wheels or robotic arms. When digital, it can be the ability to create or modify files, write reports, perform data analysis, etc.
- Brains: made of the ever increasingly complex AI tools built from neural networks, these are the decision centers of the AI agents.
- Database: this is the knowledge center with facts, training data, and human correction that allow the “brain” to make the right decision.
Source: Thomas Latterner
The combination of these components allows AI agents to have memory and a persona built on the foundations of a specialized LLM.
Memory is a very important part of the AI here and a radical improvement on previous bots. This is because lack of memory is the source of most complaints about chatbots and other similar systems: this is why most bots get stuck in loops of reasoning, why they fail to remember previous information already given, etc.
Types Of AI Agents
Besides the physical vs digital divide, there are other ways to categorize AI agents:
- One way is to consider if the agents interact with humans, or work in the background.
- Another way is to consider using a single agent for a given task or multiple agents talking to each other to perform a more elaborate job, with each agent having its own model and sharing data or not with the others.
- When using multiple agents, a hierarchy can be established, with one or several agents in charge of coordinating and “ordering around” the lower-level agents.
- Another possibility for AI agent categorization is to consider the end goals and complexity.
- Goal-based AI agents will focus on an end result, and adapt their behavior or actions until said goal is achieved. For example, a warehouse AI will give instructions for moving a package until it reaches its destination.
- Utility- based agents also focus on the goal, but also on the best way to reach it. So, for example, a self-driving car will go from point A to point B, but also take into account safety, and eventually fuel efficiency, time of travel, type of roads, etc.
Source: Ampcome
Why Use AI Agents Instead Of Generalist AIs?
There are many reasons why the AI industry is turning toward AI agents over universal models.
The first one is a matter of technical complexity and feasibility. A full human-like general intelligence (AGI) is still out of reach as far as we know.
However, it seems much more realistic to create, for example, a dedicated AI agent able to drive a car like a human while having none of the other human capacities of reasoning in other matters.
Another matter is one of efficiency. A model used to drive a car does not need to be extremely good at talking, walking, doing web searches, calculating, etc. So many individual AI agents for each task make a lot more sense than trying to deploy an all-purpose AI/robotic system, as often displayed in science fiction.
Lastly, costs are a serious issue for all AI projects. This is true of the costs of training and the thousands or even millions of GPUs it requires. But this is also true for the costs of operating the AI in computing hardware as well as the energy consumed. So more specialized, simpler AI agents, that also perform better at their job are to be preferred.
From Specialized To Generalist AI Agents?
For narrow and repeated tasks, very narrow AI agents are probably best. However, to fully grasp the benefits of the AI revolution, slightly more competent systems will be required. Like, for example, an AI is not only able to automatically generate a list of equipment needing maintenance but also schedule technicians to do said maintenance and manage their associated timesheets, salaries, etc.
In some applications, this might be a required step in truly helping human workers, as the tasks and analysis cannot fully be separated from each other.
For example, an AI performing a diagnosis will need to be able to analyze medical images, understand a text or voice describing the symptoms, integrate medical test results and patient’s history, find the relevant scientific literature and medical protocol, etc.
Source: Nature
Such generalist, but application-specific AIs are likely to be built through multi-agent constructs, with each sub-element excelling at one task, while an overseer AI will integrate the individual agents’ output into a cohesive whole.
Application-specific use cases is not the only possibility. For example, multi-agent AIs using agents from very different fields could be useful to make new scientific discoveries, by bringing together different sets of data.
AI Agent Applications
While it is likely that many applications are not even yet understood the way no one could visualize the modern Internet in 1995, a few activities are already ready for applications of AI agents:
- Customer services: From online chats to taking orders at restaurants, the relatively straightforward process of most customers’ requests makes it easier to automate. So far chatbots have been insufficient, but smarter AI agents will likely replace a lot of these jobs, with a few humans in the back to deal with the most complicated requests.
- Scientific research: This will include analyzing very complex and large datasets and browsing thousands of scientific publications, including in non-English languages. It will also cover AIs specialized in very technical tasks, like predicting protein folding, material atomic composition, etc.
- Websites & Marketing: A lot of online marketing and ads of today are already partially handled with templates and automated optimization. As more and more customized experienced are appreciated by customers, flexible AI agents are likely to soon perform as well as many humans in this field.
- Translation and Law: Many human tasks rely on having a unique set of knowledge about a very specific topic with a lot to know about, more than the topic being especially hard to analyze. The ability of AI to comb through massive amounts of data can help.
- However, the risks of “hallucinations” are very high, especially as the client is unlikely to be able to spot a mistake, so only an ultra-reliable agent will be able to perform these tasks.
- Real-time translation and voice-to-voice, especially for noncritical situations like tourism, are likely to become a standard expectation for our smartphones.
- Arts: Probably one of the most controversial functions of AI is the idea of putting out of work thousands of musicians, painters, graphic designers, etc., which does not sit well with many people. However, it could also empower much smaller teams, or even single individuals, to compete with much larger companies in making movies, video games, books, etc.
- Healthcare: Already now, Elon Musk is recommending its followers to use X’s AI, Grok, for secondary medical opinion. In the long run, it is likely that dedicated medical AI will be of assistance to doctors for analysis medical data and suggesting treatments.
- As robotic surgery becomes more commonplace, we could also imagine a future where AI surgeons can perform some surgical acts without human assistance.
- Security: this extends from local security to police and even the military. AIs can be excellent in the role of threat detection, and even target identification. However, for now, it is pretty much taboo to give them too much autonomy in this sector, especially for any lethal decisions.
- Logistics and transportation: already in increasing use in warehouses, robots, and drones are more and more likely to take over the jobs of delivery and logistics of packages and the overall supply chain as they get smarter and more able to handle real-world obstacles.
- And, of course, self-driving cars and trucks could be an even bigger revolution, completely changing how we handle mobility and potentially even making personal car ownership an oddity.
- Finance: “Algos” are already a big part of today’s financial markets, so we should expect smarter AIs to get even more involved. Customized AI agents could also become prevalent in evaluating insurance cases, loan applications, etc.
- Manufacturing: The trends of 3D printing, CNC machines, and other new tools for flexible production have made modern factories a lot more polyvalent than the old assembly line. AI agents embodied in industrial and humanoid robots could push this trend further.
Legality, Regulation & Ethics
Responsibilities
In any discussion regarding AIs, the issue of responsible handling of the technology is hotly debated.
On one side, too much regulation would hinder progress, and likely just hand over the most advanced AIs to more flexible jurisdictions. In the context of an AI tech race between the USA and China, it is clear that this is not a desirable outcome for any side.
On the other side, no one wants out-of-control AI without any responsibility.
So, a clear legal framework will need to be determined. For example, if a self-driving car crashes, is the provider of the AI agent responsible? And the more autonomy AI agents are given, the more their decisions could impact real people and turn out to be expensive.
This also covers the issue of misuse of AI, like, for example, stealing identities, committing fraud, etc.
In many cases, it is not even clear which agency or authorities should regulate AI. Should it be a dedicated specialized body? Or should AI in finance be regulated by the SEC, airborne drones, the FAA, etc?
These are more legislative and regulatory questions, but as those can often lag years behind technological progress, it is likely urgent that some of the most pressing questions regarding AI agents’ regulatory framework are answered soon.
Jobs & Inequality
An often feared effect of AI development is the growth of mass unemployment, as AI replaces more and more jobs much quicker than people can retrain or new jobs are created.
In theory, this should be a step on the path to a Utopian, post-scarcity civilization. In practice, this could push millions of people into poverty before we get to that point. And contrary to previous waves of automation, AI could replace very qualified knowledge workers.
Source: Intellipoint
The danger of monopolies or deepening wealth inequality is also a serious one, as this has historically been proven to be dangerous and destabilizing for society at large.
Ethics
What can and cannot be given to manage by AI. This is an increasingly pressing question for any tasks going beyond moving pallets in a warehouse or auto-answering emails.
The issue is even more pressing when there is a temptation to use AI for targeting systems in military drones. Including in Ukraine, with a “Rush for AI-Enabled Drones on Ukrainian Battlefields”.
Reuters reported that the race for AI-enabled drones is “taking warfare into uncharted territory as combatants race to gain a technological edge in battle.” In Ukraine, AI drone development focuses on three key areas: target identification, terrain mapping for navigation, and the creation of interconnected “swarms” of drones.
One company, Swarmer, is building software to network drones, allowing decisions to be executed instantly across a swarm with minimal human input.
Do we really want to give AIs this sort of capacity? But do we really want only “the enemy” to have it?
These issues are likely something that need to be discussed and decided internationally. They are also a topic the AI industry should not shy away from.
AI Agents Already In Service
OpenAI Agents
As a long-time leader in AI, it is not surprising to see OpenAI having several solid GPT-based agents. The company is providing developers with dedicated tools to develop AI agents, including the multi-agent Open AI SDK (Standard Development Kit).
OpenAI model seems to mostly be focused on creating better versions of GPT and other LLMs and then having them as a base used for separate agent development, counting on its leading position to generate the demand for agents using GPT.
Google (GOOGL -0.88%) has been present in AI for a long time through its DeepMind model. But it is with Gemini 2.0 that it embraced the “agentic era”.
Google is also acutely aware of the potential threat of LLM AIs to classical search, still to this day the source of 90% of the company’s revenues.
So it used Gemini 2.0 to create AI Overviews, an additional search result to tackle more complex topics and multi-step questions, including mathematics and coding.
It also created Jules, a coding assistant AI agent, and Genie 2, an AI model to create playable 3D worlds.
Google also aims to stay on the edge through hardware research, notably its Trillium TPUs (Tensor Processing Units).
Manus
Butterfly Effect, a Chinese startup, released Manus in March 2025, claiming it to be the first general AI agent capable of acting autonomously.
While some saw in it a “second DeepSeek moment” for AI agents, where China is taking the lead, the situation is less clear than with DeepSeek’s remarkably more computing (and financially) efficient approach.
Manus seems a bit slower, more prone to crash, but also providing more detailed responses than ChatGPT. Nevertheless, it seems that general AI agents might be possible sooner than expected after all, even if they might not be ideal for all situations.
Alibaba
An e-commerce leader under pressure from competitor platforms like Temu, or even TikTok, Alibaba is clawing back its position as a tech leader with progress in AI.
Notably, it open-sourced its QwQ-32B model in early March 2025, claiming that with one-fifth of the parameters of DeepSeek-R1, it is designed for efficiency.
Also, in March 2025, Alibaba released a new version of Quark, an AI assistant/agent powered by the company AI Qwen, combining deep thinking and generative AI. Before this AI revamp, Quark already had 200 million users when it was a search engine.
AI Agent Company
Alibaba
Alibaba Group Holding Limited (BABA -4.36%)
More known in the West for its e-commerce platform and as a supplier of cheap materials, parts, and consumer goods, Alibaba is also a massive tech company in China, leading in AI and cloud computing.
Notably, Alibaba controls 36% of the cloud market in China, well ahead of all its competitors.
Source: Jeff Townson
Maybe more importantly, it is already offering six new DeepSeek AI models, the open-source AI that has rocked the world by suddenly outperforming most American AI models for a tiny fraction of the costs in both development and on a per-use basis.
Alibaba also has its own AI model, Qwen, and claims Qwen 2.5 is even better than Deep Seek V3.
“Qwen 2.5-Max outperforms … almost across the board GPT-4o, DeepSeek-V3 and Llama-3.1-405B,”
Alibaba’s Cloud Unit
Overall, besides its growth in cloud and AI, Alibaba remains a giant of e-commerce in China, with Taobao & Tmall only slightly down from their 29% share of global online sales in 2019.
Source: Forbes
The recent AI progress has changed how Alibaba is seen. From a legacy e-commerce position under pressure and dominant cloud sales (but also under pressure from competitors), it has gone back to leading China’s tech innovation.
Quark is now the extra weapon Alibaba is deploying to take over the Chinese AI assistant market, having built the terrain by deploying it as an AI search engine first and gathering 200 million users.
So, despite considering its relatively low stock price, triggered by years of tech crackdown in China and concerns about investing in the country, Alibaba could be an opportunity for investors willing to bet on China taking the lead in the AI race.
(You can also read our dedicated report focused on Alibaba for more details).