The Rise of AI Agents
By 2026, advancements in artificial intelligence entered a new phase. Just a few years earlier, generative AIs, such as ChatGPT, began to solidify their role in society as entities that “answer questions.” Now, however, AI agents are increasingly taking centre stage in development and business domains.
To succinctly summarise, the distinction between generative AI and AI agents lies in their functionalities. Generative AI responds to inquiries and creates content, while AI agents operate autonomously to achieve specific objectives. For instance, if tasked with planning a business trip, a conventional generative AI would merely propose transport options like trains or flights. In contrast, an AI agent would handle the entire process automatically—searching for seat availability, making reservations, adding events to calendars, and setting reminders.
The notion of an “agent” in AI research is not a recent development; it has been recognised since the 1990s. However, it wasn’t until around 2023 that modern AI agents began to attract significant attention.
Evolution and Capabilities of AI Agents
Initially, AI agents were limited to demonstrating capabilities at the research level or executing basic automation. However, in recent years, the inferencing capabilities of large language models (LLMs) have improved dramatically. Standards for integration with external tools have been established, such as the Model Context Protocol (MCP) introduced by Anthropic in 2024. This progress has enabled AI agents to reach a level where they can be integrated into real business operations. Implementations are now being observed across various sectors, including patent work in manufacturing, sales support in financial institutions, and operations in call centres.
So, why should businesses and the general public focus on AI agents? In essence, they are defined as “an AI system that, upon receiving a goal, plans and executes the task autonomously, completing it and verifying the results without the need for constant human instructions.” This behaviour can be summarised by the following characteristics:
- Autonomy: The AI determines and executes the next action independently, without human instruction at each step.
- Goal Orientation: It develops and implements a multi-step plan to achieve a specified objective.
- Environmental Adaptation: It selects the appropriate tools based on the situation and corrects its course if errors occur during the process.
The essential characteristic of AI agents is that they complete tasks by repeating the cycle of goal-setting, planning, execution, and verification.
Contrasting AI Technologies
The differences between generative AI and AI agents can be summarised as follows:
| Generative AI | AI Agents |
|---|---|
| Principle of Operation: Question → Answer | Goal Setting → Autonomous Execution |
| Interaction: Questions and Answers | Continuous Task Execution |
| Tool Usage: May include functions like search, file management, and access to external applications depending on the product. | Selects necessary tools from available options based on goals and work progress. |
| Specific Example: Text generation; summarisation | Automated entire process (e.g., from market research report creation to submission). |
It is also crucial to understand the differences between AI agents and other existing technologies that are often confused with them:
- Chatbot: A system that responds to user queries based on predefined conversation patterns and FAQs. It can only manage a limited number of issues.
- RPA (Robotic Process Automation): A software robot that performs routine tasks automatically by following predetermined procedures. If procedures change, the script must be modified each time.
- Workflow Automation: A system that automatically executes the entire business workflow (including approvals, notifications, and branching) based on specified conditions.
These systems primarily operate based on predefined conversation patterns and processing procedures. In contrast, AI agents can leverage the reasoning capabilities of LLMs to make instant decisions and adjust plans as necessary, even in unexpected situations.
Practical Applications of AI Agents
To illustrate, let us consider a request such as: “Prepare a business trip.”
- Generative AI: Merely presents options for high-speed trains and flights. Reservations and schedule confirmations are ultimately managed by the individual.
- AI Agent: Undertakes a series of actions: searches for availability → selects and reserves the best flight → adds it to the calendar → sets a reminder for the day prior. However, for operations that are difficult to cancel, such as booking and payment, it is common to design the system to require user approval before execution.
As of 2023, AI agents faced challenges in consistently executing complex, multi-step processes, and their tool integration was limited. However, thanks to the dissemination of common standards like the MCP and improvements in LLM inferencing accuracy, they can now be tasked with more complex and time-consuming assignments.
Core Technologies Behind AI Agents
The technologies supporting AI agents can be broadly classified into three elements:
- LLM (Large Language Model): Acts as the “brain” that understands the situation, deduces the next course of action, and plans accordingly.
- Tool Integration: Mechanisms that allow interaction with the real world and systems as if they were “hands and feet”; for example: web searches, file operations, and API calls to external services.
- Loop Structure: A structure that repeats the cycle of “Execute → Evaluate → Correct,” where the results of execution are assessed, and if they do not work, corrections are made and execution is repeated.
The combination of these three elements enables the AI agent to perform a series of actions: “Receive a goal → Create a plan → Execute it → Receive results as feedback → Review the plan if necessary.”
Design Principles for AI Agents
Several frameworks are frequently used as design principles in the research and implementation of AI agents:
- ReAct (Reasoning and Action): This method involves alternating between “reasoning” and “action” phases, allowing the AI to verbalise its reasoning process and decide on actions. A key feature is that the results of actions are easily reflected in subsequent reasoning.
- Plan and Execute: This approach entails planning the entire task first and then executing each step in order. While ReAct involves “thinking and acting” sequentially, this method can be described as initially creating an overview.
- Multi-agent: A configuration where multiple AI agents with different roles collaborate to accomplish the same task. For instance, dividing tasks among agents (such as “researcher,” “writer,” and “reviewer”) simplifies the management of complex tasks that a single agent could not handle.
The “memory” mechanism is essential for AI agents to perform tasks continuously:
- Short-term Memory: Memory retained only during a single session (a set of dialogues or tasks). It is used to determine the next action based on the progression of the conversation and the outcomes of previous actions.
- Long-term Memory: Memory referred to between sessions. Typical examples include systems using RAG (retrieval-augmented generation) and vector databases. The latter converts text, images, etc., into numerical data (known as vectors) and stores them, allowing for high-speed retrieval of similar meaning information, to recover past interactions and external knowledge when needed.
Short-term memory alone is insufficient, as everything is forgotten once a session ends. Thus, the design of long-term memory is crucial for the practicality of an agent when considering continuous use.
Emerging Tools and Their Applications
Starting in 2026, the number of tools related to AI agents began to rapidly increase. The main categories can be organised as follows:
- Microsoft 365 Copilot: An agent function.
- Google Vertex AI Agent Builder: Now part of the Gemini Enterprise Agent Platform.
- Salesforce Agentforce: A powerful agent solution.
- Dify, LangChain / LangGraph, Devin, Cursor: Various tools with specific functionalities.
- ChatGPT Work: Following the path of Operator, Deep Research, and ChatGPT Agent, it incorporates Codex technology and model GPT-