AI Agent

An AI agent is a software system that uses an AI model, today usually a large language model (LLM), to pursue a goal over multiple steps. It decides what to do next, takes actions through tools, observes the results and continues until the task is done or it needs a human. The model directs the process itself instead of following a fixed script written in advance.

How an AI Agent Works

Definitions from model providers are close to each other. OpenAI's guide to building agents describes them as systems that independently accomplish tasks on a user's behalf, built from three components: a model for reasoning and decisions, tools for taking action, and instructions that set behavior and guardrails. Anthropic defines agents as systems where LLMs dynamically direct their own processes and tool usage, keeping control over how they accomplish a task.

At the core is a loop:

1. Receive a goal, from a person or another system.
2. Decide the next step, often after making a short plan.
3. Act through tool calling: search, read a file, query a database, run a test, send a message.
4. Observe the result. Feedback from the environment, such as a failing test or an empty search result, shapes the next decision.
5. Repeat or stop. The agent ends when the goal is met, when it hits a limit such as a maximum number of steps, or when it reaches a checkpoint that needs human approval.

Around that loop sit supporting parts. Instructions and guardrails define what the agent may and may not do. Agent memory carries useful information across steps and sessions. Connections to external systems are increasingly standardized through the Model Context Protocol (MCP), so one agent can use many tools without custom integration code for each.

Some systems split the work across several agents. An orchestrator plans and delegates, and subagents handle focused parts such as research, implementation or review, each in its own context. This keeps every agent's context narrow and lets a fresh agent check another's output, at the cost of more model calls and more coordination to get right.

Why AI Agents Matter

Agents handle work where the steps cannot be predicted in advance: fixing a bug whose cause is unknown, researching a market question, resolving a support ticket that needs several lookups. That is a different job from answering a single question. For product teams, the most visible example is the AI coding agent, which reads a codebase, edits files and runs tests on its own.

The trade-offs are real. Anthropic notes that agents bring higher cost and the potential for compounding errors, since a wrong early step can steer every later one. That is why the same guidance recommends starting with the simplest solution that works, adding agent autonomy only when a task needs it, and testing agents in sandboxed environments with clear guardrails. Human checkpoints, permission limits and stopping conditions are part of the design, not optional extras. Evaluation changes too: instead of grading a single answer, teams have to judge whole task runs, including the steps the agent took to get there.

AI Agent Example

A customer writes in: "My refund never arrived." A support agent built on an LLM looks up the order with one tool, checks the payment provider with another and finds the refund was issued two days ago. It drafts a reply explaining the usual bank processing time. If the refund had failed instead, it would hand the case to a human with a summary of what it checked. The steps differ from ticket to ticket, which is what makes this an agent rather than a script.

AI Agent vs. Agentic Workflow and Chatbot

Anthropic uses "agentic systems" as the umbrella and separates two kinds. In a workflow, LLMs and tools are orchestrated through code paths the developer defines in advance. In an agent, the model decides the path. The term agentic workflow is used more loosely by others, so check which meaning a source intends.

OpenAI's guide makes the same boundary explicit: simple chatbots, single-turn LLM calls and classifiers are not agents, because they do not run a workflow on their own.