An AI agent is a software system that perceives its environment, makes decisions, and takes actions to achieve a specified goal with some degree of autonomy. At its core, an agent receives observations (text, data, sensor readings), processes them using an AI model, selects an action from a defined set, executes that action, and uses the resulting feedback to inform its next step. This perception-decision-action loop is what distinguishes an agent from a passive prediction model that simply returns an output and stops.

Think of it like a human employee given a recurring job: they observe what is happening, decide what to do next, act, and adjust based on whether things went as expected—rather than waiting to be told what to do at every moment.

A minimal agent using a language model as its reasoning core might work as follows:

def agent_step(observation, memory, tools):
    prompt = build_prompt(observation, memory)
    action = llm.complete(prompt)          # model decides what to do
    result = tools[action.name](**action.args)  # execute the action
    memory.append({"action": action, "result": result})
    return result

Agents range from simple rule-based programs to LLM-driven systems that can browse the web, run code, and call third-party APIs in sequence.

AI agents are closely related to Agentic AI (the broader paradigm), Foundation Model capabilities, Chain of Thought reasoning, Prompt Engineering for structured instructions, and RAG for grounding agents in up-to-date knowledge. AI Alignment research pays particular attention to agent behavior because autonomous action in the real world raises accountability concerns.

You will encounter the term in product documentation for tools like OpenAI Assistants, Anthropic's tool-use API, and frameworks such as LangChain and AutoGen, as well as in AI Safety discussions about multi-agent systems.