An AI agent is a software system built around a large language model (LLM) that can pursue goals on its own rather than only answering one prompt at a time. It observes input, reasons about what to do next, and takes actions, such as calling an API, searching the web, running code, or writing to a database, then evaluates the result and decides its next move. This loop of perceive, reason, and act is what separates an agent from a plain chatbot.
How an AI agent works
Most AI agents combine three ingredients: a foundation model as the reasoning core, a set of tools the model is allowed to invoke, and a control loop that keeps the model working until the task is done. When given a goal, the LLM typically produces a plan, selects a tool, observes the tool's output, and updates its plan. This pattern is often called ReAct, short for reasoning plus acting.
For example, an agent told to "summarize this week's customer complaints" might list recent files, read each one, extract themes with a language model call, and write a report to a document, all without a human choosing each step. Many agents also keep short or long-term memory so they can carry context across turns and recover from errors. Frameworks like LangGraph, the OpenAI Agents SDK, and Anthropic's tool-use API expose these primitives to developers.
Why it matters
AI agents turn language models from passive responders into systems that can complete real workflows, which is why the term "agentic AI" has spread quickly across enterprise software. They are used for research assistance, coding copilots that open pull requests, customer support bots that look up account data, and operations tasks like scheduling or data entry. Because agents can chain actions and use tools, they can also fail in new ways, through wrong tool calls, infinite loops, or prompt injection from untrusted content, so they are usually deployed with guardrails, human review, and limited permissions.
Key types
- Single-step tool users: models that make one or two tool calls to answer a question, such as a chatbot that searches the web once.
- Multi-step task agents: systems that plan and execute several actions in sequence, like a research agent that reads, summarizes, and writes a report.
- Multi-agent systems: setups where specialized agents hand off work to each other, for example a planner, a coder, and a reviewer collaborating on a software task.
- Computer-use and embodied agents: agents that drive a browser, control a desktop, or operate robots by translating goals into UI or physical actions.
As foundation models improve at structured reasoning and tool use, AI agents are shifting from demos to production systems that handle end-to-end work, though reliability, evaluation, and safety remain the hard parts.