An AI agent, in the context of modern generative AI, is a system built around a Large language model that can plan a sequence of steps, take actions through external Tool use (function calling), observe the results of those actions, and iterate toward a goal, rather than simply producing one response to one prompt. The defining feature that distinguishes an agent from a plain chatbot is a loop: the model's output is not necessarily the final answer shown to a user but can instead be a decision about what to do next, which is executed, fed back into the model's Context window, and used to decide the following step, continuing until the task is complete or a limit is reached.
The term predates large language models, having long been used in artificial intelligence and robotics to describe any system that perceives an environment and acts on it, including simple rule-based agents and Reinforcement learning agents such as AlphaGo. Its current usage, tied specifically to LLM-based systems that call tools and pursue multi-step tasks, became widespread from 2023 onward, following early experiments such as AutoGPT, and accelerated sharply through 2024 and 2025 as models improved at reliably choosing and sequencing tool calls, a period widely described in industry commentary as the "agentic" wave of the broader AI boom (2022-present).
Components
A typical agent combines several pieces: a base language model that does the reasoning and decision-making, a set of tools it can invoke, such as a code interpreter, a web search function, or an API, exposed through structured tool-calling; a memory or state mechanism that tracks what has happened so far in the task, often simply the accumulated context, though some systems add external storage for longer tasks; and a control loop that decides when the task is finished or when to ask a human for input. Many agent frameworks also incorporate Retrieval-augmented generation so the agent can look up relevant documents mid-task, and some support Computer use (AI), allowing the agent to operate a graphical interface such as a web browser directly rather than through a dedicated API.
Standardization
As the number of tools an agent might connect to grew, connecting each tool to each model with custom integration code became a scaling problem. Anthropic introduced the Model Context Protocol in November 2024 as an open standard for connecting language models to external tools and data sources in a uniform way, and it was adopted broadly across the industry through 2025, including by competing labs, as a common interface layer for agentic tool use.
Applications
Agentic systems have been applied to software engineering, where tools such as GitHub Copilot, Cursor, and Claude Code evolved from single-turn code completion into agents that can read a codebase, run tests, and make multi-file edits autonomously; to research and information gathering, where an agent issues a series of searches and synthesizes findings; and to browser and computer automation, where an agent operates software the way a human user would. Benchmarks such as SWE-bench were adopted specifically to measure agentic coding performance, since they require an end-to-end sequence of file edits and test runs rather than a single code snippet.
Reliability and criticism
Agent reliability remains a central open problem: small errors early in a long task chain can compound, tool calls can fail or be misused, and agents can pursue a plausible but incorrect plan for many steps before an error becomes evident, sometimes called "agent drift." Concerns about agents acting on untrusted or adversarial input have made Prompt injection a significant safety issue specific to agentic systems, since an agent with access to real tools and data can, in principle, be manipulated into taking harmful real-world actions rather than merely generating undesirable text, a risk category with no direct analogue in a plain chatbot.