AI Agent Part 1 : What is AI Agent?

From LLMs That Answer to Systems That Act

According to the Stanford AI Index Report 2024, 63% of international survey respondents said they knew about ChatGPT, and about half of them reported using it at least once a week [1]. Behind ChatGPT’s widespread adoption lies its accessibility — it can be used easily through natural language conversation alone — and its high performance, which is strong enough to be applied directly to a wide range of real-world tasks such as writing, document summarization, and code generation.

It isn’t just ChatGPT: as various large language models (LLMs) such as Claude and Gemini have emerged one after another with increasingly sophisticated performance, they have grown beyond simply carrying out individual tasks to synthesizing and reasoning over multiple pieces of information in order to solve more complex problems [2,3].

Figure 1. Emergence and performance trends of major LLMs [4]

Recent advances in LLMs are not limited to improvements in the models themselves — they are also expanding through new ways of combining LLMs with external tools. A system in which the LLM acts as the “brain,” judging what a request requires, while external tools such as search and calculation serve as the “hands and feet” that carry out the work — flexibly adjusting its behavior based on the results — is called an “LLM-based AI agent” (hereafter, “AI agent”) [5,6].

Our research team is studying how to use AI agents to assess the impact a security patch has on an entire system. Because the effect of the same security patch can vary depending on system configuration and the operating environment, simply reading or summarizing the patch notes isn’t enough. Research on assessing security-patch impact — which requires checking related information and system status together, and continuing further impact investigation based on intermediate results — is well suited to the proactive, flexible way AI agents operate, as described above.

So what exactly is an AI agent? Does it simply mean a system in which external tools are connected to an LLM? Or does it require an operating structure distinct from how LLMs are typically used?

The scope and definition of AI agents vary somewhat across sources, but synthesizing a range of research and practical guides reveals recurring, common elements [7–10]. In this post, we’ll use those common elements to lay out the basic concept of an AI agent.

Basic Concepts of AI Agent

The Starting Point of AI Agent: Observation and Action

While today’s AI agents rely on LLMs as a core component, the concept of an “agent” itself did not first appear alongside LLMs. In the field of artificial intelligence, an agent has long been described as a system that observes its surrounding environment and selects an action based on what it observes.

For example, a robot checks its surroundings through cameras or sensors and chooses a direction to move. Game-playing AI checks the current state of the board and decides its next move. A thermostat checks the current temperature and then turns heating or cooling on or off.

Although their specific functions differ, they all share the same cyclical structure shown in the Figure below.

Figure 2. Basic structure of an agent that observes its environment and acts

<It observes and acts according to a goal>

The AI agents discussed in this post start from this same basic structure of observing the environment and selecting an action. Combined with an LLM’s ability to understand natural language and exercise judgment, together with integration with external tools, this structure lets an agent determine what tasks are needed to pursue a complex goal (objective) given in natural language, and carry them out over multiple steps.

A goal isn’t used only when the AI agent first interprets a request — it also serves as the standard for judging the next action and whether a task has been completed throughout the whole process. The AI agent determines what is needed to achieve the goal given the current situation and chooses its next action accordingly. It also keeps checking the results of its actions to judge whether the goal has been achieved or whether further work remains.

AI agents obtain the information they need for a task from the user and from the environment. Observation includes not only the user’s request but also information newly gathered during the task — such as file contents, search results, and program execution results or error messages.

<State, Memory, and Context>

For an AI agent to carry out a goal, it must manage its progress so far along with the information it needs. Task state, memory, and context each play a different role here.

Task state refers to how far the AI agent has progressed — the tasks performed so far and their results, the current stage, and what remains to be done. Memory is a structure that stores information gathered during a task, or past records, so they can be reused later. Context is the information actually provided to the LLM when the AI agent judges its next action. It can include the user’s request, the current task state, the result of the most recent execution, and relevant information retrieved from memory.

In short: task state represents “how far things have progressed,” memory represents “what is being kept,” and context represents “what the LLM is looking at and judging right now.”

<Choosing actions based on the LLM’s judgment and policy>

An AI agent judges the next action it needs to achieve its goal based on a context built by reflecting current observations and task state. In this process, the LLM plays the central role of understanding the user’s request and the current task situation, and selecting one action from among the available options.

However, the choice of action isn’t determined by the LLM’s judgment alone. Policies — such as instructions given to the system, task procedures, the tools and permissions available, approval conditions, and safety rules — are factored in as well.

For example, a system can be configured so that if information is insufficient it looks for additional data; before carrying out a risky action it asks the user for approval; and once enough evidence has been gathered it wraps up the results and ends the task.

<Planning and coordinating the workflow>

When carrying out a complex goal, planning may be used to break the goal into several sub-tasks and decide the order in which to perform them. For example, an AI agent reviewing the impact of a security patch might structure its work as: analyze the patch announcement, investigate related conflict cases, compare them against the current system configuration, and, if necessary, verify them in a test environment.

Such a plan doesn’t need to stay fixed from start to finish. If an unexpected conflict case turns up in search results, or an error occurs during a test run, the AI agent can revise the plan or return to an earlier step to incorporate the new information.

That said, not every AI agent has an explicit plan or a separate planning module. For simple tasks, rather than drawing up an entire plan in advance, an agent can also operate by choosing its next action one step at a time, based on the current situation and the most recent result.

Meanwhile, even when an LLM is connected to multiple tools, if every step and execution order is fixed in advance, this is sometimes distinguished from an agent and called an “LLM workflow.” What sets an AI agent apart is that it can dynamically choose its next step, or which tool to use, based on the current task state and intermediate results. In practice, however, predefined procedures and an AI agent’s dynamic judgment are often used together within the same system [7].

<Actions and tool use>

Actions an AI agent can choose include responding to the user, requesting additional information, revising the plan, and pausing or ending the task. When needed, tool use — such as reading or saving files, searching the web, executing code, or querying a database — can also be selected as an action.

When a tool runs, it returns results such as file contents, search results, program output, or error messages. These execution results become new observations: the AI agent reflects them in its task state and memory, and then builds the context it needs for its next judgment.

Based on the execution results, the AI agent checks whether the goal has been achieved and decides whether to continue or end the task. This cycle continues until the goal is achieved or further human judgment is required.

Figure 3. The operating structure of an AI agent repeating observation and action

As we’ve seen, an AI agent isn’t built from a single function. Goals and observation, task state and memory, planning, and actions and tool use are all connected, operating together as one system. Various studies on AI agents likewise treat memory, planning, and actions and tool use as key components and capabilities in particular [9].

Bringing these common characteristics together, an AI agent can be defined in a single sentence as follows:

An AI agent is an AI system that observes information from the user and the environment in order to achieve a given goal, and selects and executes its next action based on the current task state and intermediate results. It uses external tools when necessary, re-observes the results of its actions to adjust subsequent work, and, for complex tasks, may draw up a plan or revise that plan in light of new results.

Conclusion

The concept of an “agent” existed before LLMs, but as LLMs’ natural language understanding and judgment capabilities combined with integration into external tools, a wide variety of AI agents emerged. Along the way, the scope and components of AI agents have also come to be described somewhat differently across different sources. In this post, we focused on the elements that consistently appear across these descriptions to lay out the basic concept of an AI agent.

In the next post, we’ll look at how these concepts are implemented in an actual system. We’ll introduce what information the security-patch-impact AI agent our team is developing observes, what tools it uses, and how these elements connect within a single workflow — we hope you’ll look forward to it.

References

[1] Stanford Institute for Human-Centered Artificial Intelligence, “The AI Index 2024 Annual Report”, Chapter 9: Public Opinion, Stanford University, 2024.
[2] Stanford Institute for Human-Centered Artificial Intelligence, “The AI Index 2025 Annual Report”, Chapter 2: Technical Performance, Stanford University, 2025.
[3] Stanford Institute for Human-Centered Artificial Intelligence, “The AI Index 2026 Annual Report”, Chapter 2: Technical Performance, Stanford University, 2026.
[4] Artificial Analysis, State of AI: China, Q1 2025
[5] T. Schick et al., “Toolformer: Language Models Can Teach Themselves to Use Tools,” Advances in Neural Information Processing Systems, vol. 36, 2023.
[6] S. Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models,” International Conference on Learning Representations, 2023.
[7] Anthropic, “Building Effective Agents,” 2024.
[8] S. J. Russell and P. Norvig, “Artificial Intelligence: A Modern Approach”, 4th ed., Pearson, 2021.
[9] L. Wang et al., “A Survey on Large Language Model Based Autonomous Agents,” Frontiers of Computer Science, vol. 18, article 186345, 2024.
[10] OpenAI, “A Practical Guide to Building Agents”, 2025.