Some of this article's listed sources may not be reliable. Please help improve this article by looking for better, more reliable sources. Unreliable citations may be challenged and removed. (December 2025) (Learn how and when to remove this message)

Within the scope of generative artificial intelligence, AI agents (often termed compound AI systems or agentic AI) act as a type of intelligent agent capable of striving for objectives, employing tools, and executing actions with differing levels of independence. In actual use, they usually function within bounds and goals set by humans, using available tools.^[1][2]

AI agents are characterized by several core features, such as goal-oriented actions, natural language interfaces, the ability to leverage external tools, and the competence to handle multi-step tasks. Their control logic is largely powered by large language models (LLMs). These systems may also incorporate memory modules, planning capabilities, tool connections, and orchestration software to coordinate various parts of the agent.^[2][3]

There is no universally accepted definition for AI agents.^[4][5][6] NIST has identified agentic AI as a nascent field that requires standards for secure operations, compatibility, and dependable interactions with external systems.^[1]

A typical use case for AI agents is task automation, such as arranging travel itineraries based on a user's prompted request.^[7][8][9]

Major technology firms like Google, Microsoft, and Amazon Web Services have introduced platforms for deploying ready-made AI agents.^[10] Various protocols have been suggested to standardize communication between agents, including the Model Context Protocol, Gibberlink,^[11] and others. Some of these protocols also facilitate connections between agents and external software.^[12]

In December 2025, the Linux Foundation launched the Agentic AI Foundation (AAIF) to ensure that agentic AI develops in a transparent and collaborative manner.^[13][14]

The roots of AI agents can be found in research from the 1990s, with Harvard professor Milind Tambe observing that the definition of an AI agent was unclear even back then. Researcher Andrew Ng is credited with popularizing the term "agentic" among a wider audience in 2024.^[15]

Efforts have been made by researchers to construct world models^[16][17] and reinforcement learning environments^[18] for the purpose of training or assessing AI agents. For instance, video games such as Minecraft^[19] and No Man's Sky^[20], along with duplicates of corporate websites,^[21] have served as training grounds.

The Financial Times has likened the autonomy of AI agents to the SAE levels for self-driving cars, placing most current applications at level 2 or 3, with some reaching level 4 in specific niches, and level 5 remaining a theoretical concept.^[22]

Several internal design strategies for reasoning within an agent have been proposed:^[23]

• Retrieval-augmented generation
• The ReAct (Reason + Act) pattern, a repetitive process where the AI agent switches between thinking and acting, receiving feedback from the environment or tools, and using that information for subsequent steps.^[24]
• Reflexion, which employs an LLM to generate feedback on the agent's strategy and stores it in a memory buffer.
• A tool/agent registry, used to organize software functions or other agents accessible to the agent.
• One-shot model querying, a method where the model is queried a single time to generate an action plan.

Ken Huang suggested a reference architecture for AI Agents composed of seven interconnected layers, each relying on the layer below it^[25]:

• Layer 1: Foundation models - serve as the primary AI engines driving the agent.
• Layer 2: Data operations - handle the intricate data infrastructure necessary for operations, including Vector databases, data loaders, and RAG.
• Layer 3: Agent frameworks - advanced tools designed to streamline the creation and management of AI agents.
• Layer 4: Deployment and infrastructure - supply the solid technical base required to run AI agents.
• Layer 5: Evaluation and observability - concentrate on measuring the safety and effectiveness of AI agents.
• Layer 6: Security and compliance - acts as a vital protective layer ensuring agents function safely and adhere to regulations, integrating security features across the entire stack.
• Layer 7: Agent ecosystem - represents the interaction point between AI agents and real-world users or applications.

To handle complex jobs, autonomous agents are frequently combined with other agents or specific tools. These setups, known as orchestration patterns or workflows, include the following:^[26][27]

• Prompt chaining: A linear sequence where the result of one step becomes the input for the next.
• Routing: The categorization of an input to guide it to a specific downstream task or tool.
• Parallelization: Running multiple tasks at the same time.
• Sequential processing: A predetermined, linear path for tasks.
• Planner-critic: A cycle where one agent creates a plan and another reviews it to offer suggestions for improvement.

Beyond large language models (LLMs), vision-language models (VLMs) and multimodal foundation models can also serve as the foundation for agents. In September 2024, the Allen Institute for AI published an open-source vision-language model.^[28] Nvidia released a fram