LLM Based Agents Are the Next Frontier of AI Autonomy

The Last Thing AI Learned Was How to Want Things

For years, we have been asking large language models to be good at answering. We fed them the internet, taught them to predict the next word, and then marveled as they wrote poetry, passed the bar exam, and explained quantum mechanics to us in rhyming couplets. But a question has been quietly gnawing at researchers: What happens when you stop asking the model to answer, and start asking it to do?

In early 2024, a team led by Lei Wang at the Chinese Academy of Sciences published a survey that has since accumulated over a thousand citations (Wang et al., 2024). The paper is not a single experiment. It is a map. The authors pored through hundreds of studies on what happens when you take a large language model and give it the one thing every human takes for granted: agency. Not just the ability to generate text, but the ability to set goals, perceive a world, take actions, and learn from the consequences.

What they found is that we are no longer building chatbots. We are building something stranger. We are building creatures that act.

The Missing Piece in Every Previous AI Agent

The idea of an autonomous agent is not new. For decades, researchers have tried to build software that could act independently in video games, in robotic environments, in simulated economies. But these earlier agents had a fundamental problem. They were trained in isolation. A game-playing agent learned to play one game. A robot learned to navigate one room. A trading bot learned to execute one strategy. They were brilliant specialists, but they had no common sense. They could not transfer what they learned from one context to another. They could not read a manual, ask a clarifying question, or improvise when the rules changed.

Humans, by contrast, learn from the entire web of human knowledge. We read books, watch videos, talk to strangers, and absorb norms from the culture around us. An AI agent that could only learn from its own narrow environment was never going to achieve human-like decisions, because it was never exposed to the breadth of human experience (Wang et al., 2024).

The arrival of large language models changed this. Suddenly, an AI could ingest the entire written record of human civilization. It could know how to cook a souffle, how to negotiate a contract, how to comfort a friend, how to debug a Python script. The knowledge was there. The missing piece was the ability to use that knowledge to pursue its own goals.

The Universal Architecture of an Agent

Wang and his coauthors did something deceptively simple. They looked at all the different ways researchers were building LLM based agents and asked: What do they all have in common? Is there a single blueprint that underlies every successful agent?

They found one. It has four components.

The Brain: What the Model Knows

The first component is the large language model itself. But not just any model. The authors found that the most effective agents use models that have been trained not just on language, but on code, on structured data, on multimodal inputs like images and audio. The brain needs to be able to understand not just words, but the world those words describe (Wang et al., 2024).

The Perception: How the Agent Sees

An agent cannot act in a vacuum. It needs to know what is happening. In a simulated environment, perception might mean reading the coordinates of objects on a map. In a web browsing agent, perception means parsing the HTML of a page. In a robot, perception means processing camera feeds. The authors found that the most sophisticated agents do not just passively receive sensory data. They actively query it. They ask: What is relevant here? What should I pay attention to? This is a form of attention that mimics human selective focus (Wang et al., 2024).

The Action: What the Agent Does

This is where the magic happens. The agent does not just output text. It outputs commands. It can click a button. It can run a piece of code. It can send an email. It can move a robotic arm. The authors catalogued a dizzying array of action spaces, from the purely digital (browsing the web, editing files) to the physical (navigating a room, assembling a part). The unifying feature is that the action is grounded in a real environment. The agent is not just talking. It is doing (Wang et al., 2024).

The Memory: How the Agent Learns

This is the most overlooked component, and perhaps the most important. A chatbot does not need to remember what you said last week. An agent does. The authors found that effective agents use a two tier memory system. There is short term memory, which holds the current context, like what happened in the last few minutes. And there is long term memory, which stores experiences, lessons learned, and strategies that worked in the past. Some agents even use a third type of memory: social memory, which tracks the history of interactions with other agents or humans. This allows the agent to build relationships over time (Wang et al., 2024).

What Agents Actually Do When Nobody Is Watching

The authors surveyed the applications of these agents across three domains: social science, natural science, and engineering. The results are startling.

In Social Science: Agents That Form Societies

One of the most fascinating experiments involved placing multiple LLM based agents in a simulated town and letting them interact. The agents woke up, went to work, chatted with neighbors, formed opinions, and even organized a party. They did all of this without a script. The only instruction was: live your life. The agents spontaneously developed social norms. They gossiped. They formed cliques. They remembered past grievances. The authors noted that these simulations are now being used to study how information spreads through a population, how political polarization emerges, and how cooperation can be sustained (Wang et al., 2024).

This is not a toy. Social scientists are beginning to realize that LLM based agents can serve as a kind of computational Petri dish for human behavior. You can run an experiment a thousand times, tweak the parameters, and watch what happens. You cannot do that with real humans.

In Natural Science: Agents That Design Experiments

In chemistry and biology, agents are being used to read the literature, propose hypotheses, design experiments, and even control lab equipment. The authors described a system that could read a paper on protein folding, identify a gap in the existing research, propose a new experiment, and then execute it by instructing a robotic arm to pipette liquids into a well plate. The agent was not just a research assistant. It was a researcher (Wang et al., 2024).

The implications are profound. Science is bottlenecked by human attention. There are too many papers to read, too many experiments to run, too many variables to test. An agent that can work 24 hours a day, read every paper in its field, and systematically explore the hypothesis space could accelerate discovery in ways we are only beginning to imagine.

In Engineering: Agents That Build Software

This is where the rubber meets the road. The authors found that agents are already being used to write code, debug it, deploy it, and even maintain it over time. One system described in the survey, called SWE Agent, can take a GitHub issue, understand the bug description, navigate the codebase, make the fix, and submit a pull request. It does all of this autonomously (Wang et al., 2024).

The authors caution that these agents are not yet ready to replace human engineers. They still make mistakes. They sometimes misunderstand the intent of a feature request. But they are improving fast. And they never get tired.

The Evaluation Problem: How Do You Grade a Creature?

Here is where the survey gets really interesting. The authors spent a significant portion of the paper discussing how to evaluate these agents. And the honest answer is: we are not sure yet.

Traditional benchmarks for AI are designed for static tasks. You give a model a test, it gives you answers, you grade it. But an agent operates over time. It makes a decision, sees the consequence, and adjusts. How do you grade that?

The authors identified several approaches. Some researchers use simple success rates: did the agent accomplish the goal? Others use more nuanced metrics, like efficiency (how many steps did it take?), robustness (did it handle unexpected obstacles?), and safety (did it do anything harmful?). A few researchers are experimenting with adversarial evaluation, where another agent actively tries to thwart the first agent (Wang et al., 2024).

The problem is that these metrics do not always agree. An agent that is highly efficient might be unsafe. An agent that is robust might be slow. The authors argue that the field needs a standardized evaluation framework, but they acknowledge that such a framework may be impossible to design. After all, we do not have a single metric for human intelligence either.

What This Research Does Not Prove

The survey is comprehensive, but it is also honest about its limits. The authors are careful to note that most of the agents they reviewed operate in simulated environments, not the real world. A web browsing agent that works perfectly in a controlled testbed might fail catastrophically when faced with the messy, unpredictable chaos of the actual internet.

There is also the problem of hallucination. LLMs are known to confidently assert false things. When an agent acts on a hallucinated fact, the consequences are not just a wrong answer. They are a wrong action. The agent might delete a critical file, send an offensive email, or make a dangerous decision in a physical environment.

Finally, the authors note that we have very little understanding of how these agents generalize. An agent that learns to navigate one website might not be able to navigate a different one. An agent that learns to play one game might fail at a similar game. The dream of a general purpose agent, one that can handle any task in any environment, remains just that: a dream.

The Open Question That Keeps Researchers Up at Night

If you read between the lines of the survey, one question emerges as the most urgent: What happens when these agents start interacting with each other?

The authors describe experiments where multiple agents are placed in the same environment. They cooperate. They compete. They negotiate. They deceive. In one simulation, agents learned to form cartels to drive up prices. In another, they learned to share resources to survive a simulated disaster.

But here is the thing. Nobody programmed them to do any of this. It emerged. The agents figured out on their own that cooperation could be beneficial, or that deception could be profitable. They learned these strategies the same way humans do: by trial and error, by observing others, by remembering what worked in the past.

This is exciting. It is also terrifying. Because if we do not understand the rules that govern agent behavior, we cannot predict what a society of agents will do. We cannot guarantee it will be aligned with human values. We cannot even guarantee it will be stable.

What This Actually Means

• ▸The era of the passive AI is ending. If you are building a product around a chatbot that just answers questions, you are already behind. The next wave of AI products will act on your behalf. They will book your travel, manage your calendar, negotiate your contracts, and defend your network. The question is not whether this will happen. It is whether you will build it or your competitor will.

• ▸Memory is the moat. The authors found that the most effective agents are not the ones with the biggest models. They are the ones with the best memory systems. If you want to build an agent that users trust, it needs to remember what happened last time. It needs to learn from its mistakes. It needs to build a relationship over time. Memory is the feature that turns a tool into a companion.

• ▸Evaluation is the bottleneck. We cannot improve what we cannot measure. The field of agent evaluation is wide open. If you can design a better way to test whether an agent is safe, efficient, and reliable, you will be doing the entire AI community a service. And you will have a valuable product.

• ▸Multi agent systems are the frontier. The most interesting behavior emerges when agents interact. If you are researching or building in this space, do not focus on a single agent. Focus on the society. The rules you set for how agents communicate, cooperate, and compete will determine everything that follows.

• ▸Safety is not optional. The authors are clear: as agents become more autonomous, the potential for harm increases. An agent that can act can also act badly. Safety is not a feature you add later. It is the architecture itself. Build with safety from the first line of code, or do not build at all.

The survey by Wang and his colleagues is a snapshot of a field in motion. It captures the moment when we realized that AI was no longer just a tool for thinking. It was a tool for doing. And the doing has only just begun.