What Is an AI Agent? Definition, Types and Real Examples

"AI agent" now labels everything from a chat window to software that writes and ships code by itself. If you are a business owner, a student or a manager trying to pin down what the word means, the definitions you find disagree in small ways that matter.

This guide gives one plain definition, walks through how an agent works with one example, then covers its parts, types, real products, failure modes and how to test one before you trust it.

What is an AI agent?

An AI agent is a program that is given a goal rather than a single question. It uses an AI model, usually a large language model, to decide what to do next, takes that step with a tool, looks at what happened and keeps going until the goal is met, it gets stuck, or a person needs to approve something.

A large language model (LLM) is an AI model trained on huge amounts of text to predict and generate language. On its own it only produces text. What turns it into an agent is the loop around it: permission to call tools, a record of what happened and rules for when to stop.

AI agent definition: where does the term come from?

The idea of an AI "agent" comes from decades of AI research, long before chatbots. The textbook Artificial Intelligence: A Modern Approach by Stuart Russell and Peter Norvig builds the whole subject around it: the preface of the fourth edition (2021) says "the main unifying theme is the idea of an intelligent agent" and defines AI as "the study of agents that receive percepts from the environment and perform actions." Chapter 2 in the official table of contents is titled "Intelligent Agents".

A percept is whatever the agent senses at a given moment: a temperature reading, a camera frame or a customer's message. Under that definition a thermostat and a chess program are agents too; the definition says nothing about language.

One of the papers that brought language models into the story is ReAct, published in October 2022 by Shunyu Yao and colleagues, in which a model alternates between written reasoning and actions such as querying a Wikipedia API. Reason, act, observe, repeat: that is the loop nearly every agent runs today.

Put together, an agent needs four things: a goal, a model that picks the steps, tools to act with and feedback from each step. Take one away and you have something else, usually a chatbot or a fixed workflow.

How does an AI agent work?

An AI agent works by running a loop: it reads the goal, picks a step, carries it out with a tool, reads the result and decides again until it reaches a stop condition. OpenAI's guide says the "concept of a while loop is central to the functioning of an agent", and Anthropic's says an agent must get "ground truth" from its environment at each step.

Here is the loop with one example. A shop's customer writes to its support agent: "Order 4821 still hasn't arrived. Can you check, and open a ticket if something's wrong?"

1. Read the goal and the rules. The agent gets the message, its instructions ("orders not shipped after 5 working days get a ticket") and its tools: an order lookup, a ticket tool and a handoff to a person.
2. Decide the next step. The model reasons that it needs the order's status first, so it chooses the order lookup with the input "4821".
3. Act. The software around the model calls the shop's system. The model only proposes the call; the code runs it and can refuse it.
4. Observe. The result: paid, not shipped, placed eight days ago.
5. Decide again. Eight days breaks the five-day rule, so the model calls the ticket tool with a short summary and the customer's email.
6. Check the result. The tool returns ticket 5531. Had it returned an error, the agent would retry once or hand the conversation to a person.
7. Stop and report. The goal is met, so the agent tells the customer the order has not shipped and ticket 5531 is open.

OpenAI lists common exits: a tool call, a set final output, an error or "reaching a maximum number of turns"; a good agent also stops when a rule says a person must decide. A scripted chatbot would have stopped at step 1, and a fixed workflow would run the same steps whatever the order said.

What are the parts of an AI agent?

An AI agent has five parts: a model that decides, instructions that set the goal and rules, tools that let it read and change things, memory that carries context, and guardrails that limit what it can do. OpenAI's guide names model, tools and instructions as the "core components"; Anthropic's describes "an LLM enhanced with augmentations such as retrieval, tools, and memory."

A tool is a function the agent may call, described by a name, a note on when to use it, its inputs and its result. Many agents now reach tools through the Model Context Protocol (MCP), which Anthropic's Claude Code documentation calls "an open standard for connecting AI tools to external data sources."

Memory is what the agent carries forward: the current conversation and tool results (short-term) and anything stored for later sessions (long-term). A support agent also uses retrieval, a search over your own documents, so it answers from your policies rather than from the model's training.

A guardrail is a limit enforced in code or by a separate check, so the agent cannot talk its way past it: a read-only tool, a step cap, a field list that hides addresses. Instructions say what the agent should do; guardrails decide what it can do.

What is the difference between an AI workflow and an AI agent?

An AI workflow follows steps a developer wrote in advance, while an AI agent decides its own steps as it goes. Anthropic's December 2024 guide, Building effective agents, says "workflows are systems where LLMs and tools are orchestrated through predefined code paths," while agents are "systems where LLMs dynamically direct their own processes and tool usage." It calls both "agentic systems".

OpenAI's practical guide to building agents adds that apps which use a language model without letting it control the workflow, such as "simple chatbots, single-turn LLMs, or sentiment classifiers", are not agents. It recommends agents for complex judgement calls, rule sets too tangled to maintain, and work that depends on unstructured text.

The trade-off is cost against flexibility. Anthropic writes that agentic systems "often trade latency and cost for better task performance" and suit "open-ended problems where it's difficult or impossible to predict the required number of steps." Its advice: find "the simplest solution possible, and only increasing complexity when needed."

What are the types of AI agents?

AI agents are sorted two ways: the five classic types from the Russell and Norvig textbook, which describe how an agent decides, and the practical types sold in 2026, which describe the job it does.

The five classic types of agents

Section 2.4 of the textbook, "The Structure of Agents", describes five designs.

In this scheme, most of today's language-model agents work like goal-based agents: given a goal, they choose actions towards it.

The practical types of AI agents in 2026

These are the job-based types on product pages; examples follow below.

AI agent vs chatbot, AI assistant, copilot and automation: what is the difference?

The difference is who decides the next step and whether the software can act: an AI agent decides and acts, a chatbot replies, an assistant or copilot suggests while a person decides, and automation acts only along fixed steps.

An AI assistant, in Google Cloud's description, "can recommend actions but decision-making is done by the user." A copilot is an assistant inside a person's own tool, such as a help desk. Workflow automation runs a fixed chain of steps when a trigger fires, like a Zapier flow that adds each form entry to a spreadsheet. Robotic process automation (RPA) is software that replays recorded clicks and keystrokes in other applications.

The page on what an AI support agent is compares the support versions.

What are some examples of AI agents?

Here are eight real AI agents, each described from its maker's own page on 12 November 2026. PepoChat is our product; we include it where it fits and say where other tools are the better choice.

Claude Code and OpenAI Codex: coding agents

Anthropic describes Claude Code as "an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools." OpenAI's Codex runs as a command-line tool, an IDE extension and a cloud service, for tasks such as building a feature or fixing failures.

Gemini Deep Research: a research agent

Google calls Deep Research "an agentic feature in Gemini that can automatically browse up to hundreds of websites" and write a multi-page report. It first shows a research plan you can refine, a small human checkpoint.

Claude in Chrome: a browser agent

Anthropic describes Claude in Chrome as "a browser extension that allows Claude to read, click, and navigate websites alongside you", and warns that it is "enhanced with our safety classifiers but is still risky."

Salesforce Agentforce and PepoChat: customer support agents

Salesforce calls Agentforce "the AI agent platform that delivers 24/7 autonomous support at enterprise scale", and it is the better choice where customer data already lives in Salesforce. PepoChat is a support agent for smaller websites that answers from your knowledge base, runs actions and hands off to your team inbox.

ElevenLabs' ElevenAgents: voice agents

ElevenAgents, ElevenLabs' agents platform, combines speech-to-text, a language model of your choice, text-to-speech and a turn-taking model, with tools and phone lines through SIP trunking or Twilio. More in the voice agents guide.

Anthropic's Research system: a multi-agent example

In June 2025 Anthropic described the multi-agent system behind Claude's Research feature: a lead agent hands parts of a question to sub-agents that search in parallel. It reports that a lead Claude Opus 4 agent with Claude Sonnet 4 sub-agents beat a single-agent Claude Opus 4 by 90.2% on its internal research evaluation.

What are AI agents good at, and where do they fail?

AI agents are good at multi-step tasks whose path cannot be scripted but whose result can be checked. They fail through errors that build up across steps, inconsistency, cost, manipulation through text they read, and actions that cannot be undone.

Anthropic's guide gives the reasons coding and support suit agents: "code solutions are verifiable through automated tests", and in support "success can be clearly measured through user-defined resolutions." The failures are just as well documented:

• Errors compound. Anthropic wrote in 2025 that "agents are stateful and errors compound": an early mistake sends every later step the wrong way. The arithmetic below is an illustration, not a measurement.
• Inconsistency. The τ-bench paper (Yao and colleagues, June 2024), which tests agents on simulated retail and airline support, found that state-of-the-art function-calling agents such as gpt-4o succeeded on fewer than half of the tasks, and that their chance of solving a retail task in all eight of eight attempts was under 25%.
• Cost. Anthropic's 2025 post reports that agents "typically use about 4× more tokens than chat interactions", and multi-agent systems about 15 times more.
• Prompt injection. Prompt injection is text written to override an AI's instructions. OWASP's Top 10 for LLM applications (2025) ranks it first, including an indirect form hidden in pages or files the model reads.
• Irreversible actions. OpenAI's guide says "canceling user orders, authorizing large refunds, or making payments" should trigger human oversight.

Illustration: a 10-step task where each step independently succeeds 95% of the time

Chance the whole task succeeds = 0.95 x 0.95 x ... (10 times) = 0.95^10
0.95^10 = 0.599, so about 60%

The same task at 99% per step: 0.99^10 = 0.904, about 90%

So shorten chains, check results between steps and stop early when something looks wrong. The guide to chatbot permissions sorts support actions into what an agent may do alone and what needs a person.

How much autonomy should an AI agent have?

An AI agent should have the least autonomy that gets the job done, with a person approving anything costly or irreversible. K. J. Kevin Feng, David W. McDonald and Amy X. Zhang proposed five levels of autonomy for AI agents in June 2025, defined by the role the person plays.

OpenAI's guide names two triggers for bringing in a person: the agent exceeding a failure threshold, such as misunderstanding a request several times, and high-risk actions. Anthropic adds that agents can "pause for human feedback at checkpoints or when encountering blockers."

How do you evaluate an AI agent before you trust it?

You evaluate an AI agent by running it several times on real tasks with a written success criterion, reading the transcripts and checking the outcome in the real system. Anthropic's January 2026 guide, Demystifying evals for AI agents, sets out the method, and it works for buyers too.

1. Collect 20 to 50 real tasks. Anthropic says "20-50 simple tasks drawn from real failures is a great start." Take them from your inbox, not a demo script.
2. Write what success means. For support: "the right order status, no address shown, ticket filed only if late."
3. Run each task several times. Results vary between runs; Anthropic's pass^k measures "the probability that all k trials succeed", which is what a customer experiences.
4. Read the transcripts. "You won't know if your graders are working well unless you read the transcripts and grades from many trials."
5. Check the outcome, not the claim. Look in the order system or ticket queue; "done" in the chat proves nothing.
6. Try to break it. Hide an instruction in a test document, make a tool fail, ask for a refund it should refuse.
7. Re-run after every change. A fix for one case can break three others.

How can a small business use an AI agent today?

A small business gets the most from an AI agent by giving it one narrow job with a clear finish line, such as answering website questions from its own content, checking order status, booking appointments or collecting lead details, and keeping a person for everything else.

Buying usually beats building at this size, because a developer kit leaves you to own hosting, monitoring, tool security and tests. Start read-only for two weeks, read what the agent did, then add one action that changes something, with a confirmation step. Ask each vendor how it bills (per reply, resolution, seat or token), because agents make more model calls than a chat.

For website support, PepoChat is one option: one script tag installs it, and its Free plan covers 100 AI replies a month. The comparison of AI agents for customer service automation lists others and what each can do.

Common misconceptions about AI agents

"An AI agent is a robot." Most agents have no body. They act through other software: an API call, a browser click, a ticket.

"Agent means no human." Anthropic's and OpenAI's guides both build a person into the design; OpenAI calls human intervention "a critical safeguard". A good agent hands over cleanly instead of guessing.

"More autonomy is always better." Autonomy adds cost, delay and risk. Anthropic's advice is the simplest system that works.

"It learns from every conversation." Most business agents run on a fixed model. They improve when a person reads transcripts and fixes the instructions, documents or tools.

"Anything with AI in a chat window is an agent." If it cannot choose a step and use a tool, it is a chatbot. The look at agentic AI in customer service shows how far real support agents go today.

How PepoChat uses AI agents

PepoChat is an AI agent built for one job: customer support on a website. Each product question starts with a search of your knowledge base, the pages, files and Q&A pairs you add, and the agent answers from what it finds.

It acts through tools. There are 22 prebuilt actions across 13 providers, such as Shopify order status by order number, Stripe subscription checks and ticket creation in Zendesk, Freshdesk or Zoho Desk, plus custom actions against any HTTPS endpoint with up to 10 inputs, as the custom actions docs describe. Procedures are playbooks of up to 15 steps with if/then branches that the agent follows for a named situation. It books meetings through Cal.com or PepoSmart, and hands the conversation to your team inbox when it cannot answer or the visitor asks for a person.

The guardrails are part of the design: you choose which actions are enabled, saved agent tests grade scripted conversations against a success criterion you write, and no action moves money. Execution is at-least-once, so a write can very rarely run twice, which is why refunds and payments should not be wired as actions. You do not choose the AI model.

Every chat feature is on the Free plan of 100 AI replies a month, and paid plans start at $39 a month on the pricing page. You can create a free workspace with no card and test the agent on your own questions.

What to do next

• Read what an AI support agent is for the customer-service version of this definition.
• Before you switch on any action, sort your tasks with chatbot permissions: what not to allow.
• Compare reply allowances on the pricing page.
• Start a free PepoChat workspace, add your website as a knowledge source and run the seven evaluation steps above.
• Related explainers: What is agentic AI?, What is MCP? and What is prompt injection?.