What is an AI agent? Definition, types and examples

Published by 3Gen Internet Corporation, maker of OpenAssistant, Skej and AGNT. Comparisons use public documentation, without hands-on testing.

An AI agent is software that uses an AI model to work toward a goal for you and chooses its own steps and tools, instead of following a fixed script. The big vendors word it differently: Anthropic describes agents as “systems where LLMs dynamically direct their own processes and tool usage” and separates them from workflows that run on predefined code paths (Anthropic), OpenAI says they “independently accomplish tasks on your behalf” (OpenAI guide), Google Cloud says they “pursue goals and complete tasks on behalf of users” (Google Cloud), and Microsoft calls an agent “a flexible software program that uses generative AI models” (Microsoft Learn).

How an agent differs from a workflow, a chatbot and an assistant

The four sources do not share one definition, and the differences matter when a vendor calls its product an “agent”.

Workflow automation. Anthropic defines workflows as “systems where LLMs and tools are orchestrated through predefined code paths” (Anthropic). A person fixes the steps in advance. Anthropic treats workflows and agents as two kinds of “agentic systems” and recommends starting with the simplest one that works. OpenAI draws a harder line: its guide says an application that does not use the model to control how a workflow runs is not an agent (OpenAI guide). Microsoft contrasts agents with fixed-rule applications, saying agents “dynamically orchestrate workflows based on real-time context” (Microsoft Learn).

Assistant. Google Cloud treats an AI assistant as a kind of agent that can “reason and take action on the users’ behalf with their supervision” (Google Cloud). Google compares agents, assistants and bots by autonomy and says agents have the highest degree of it. An assistant that acts for you is the supervised end of that range.

Chatbot. The four sources do not give a separate chatbot definition beyond Google’s autonomy comparison. In this article, “chatbot” is our working term for a product that replies to messages in a chat window and does not act in your other tools. If a product only answers, it is a chatbot. If it also sends, books, edits or buys, the question becomes how much it decides on its own.

A concrete example of the assistant end is OpenAssistant, which is made by the company that publishes this site. Its pages describe an assistant you reach “over calls, iMessage, SMS, WhatsApp, email, or the web app” (OpenAssistant), and the help center lists Slack as a request channel too (OpenAssistant help). It has its own inbox, phone number, calendar and a sandboxed machine where it opens apps and websites (OpenAssistant). You set approval per kind of action, so the same product can ask first or act on its own (OpenAssistant help).

Seven types of AI agent, with one example each

These are editorial groupings by job. Each example is one product, picked from the vendor’s own pages, not a ranking and not a recommendation.

1. Coding agents. Anthropic’s Claude Code documentation calls it “an agentic coding tool that reads your codebase, edits files, runs commands” (Claude Code docs). Its cloud sessions keep running after you close your laptop (Claude Code on the web). The work product is code and pull requests.

2. Browser and computer-use agents. OpenAI’s help page says a cloud browser “gives ChatGPT Work its own browser on a separate computer in the cloud” (OpenAI help). By default it asks permission before opening a new website and asks for confirmation before actions that could be hard to reverse, such as a booking or a payment. The same help center says the older “ChatGPT agent is no longer available” (OpenAI help), a reminder that product names in this field change fast.

3. Research agents. OpenAI describes deep research in ChatGPT as “reasoning, researching, and synthesizing information into a documented report” (OpenAI help). The help article says ChatGPT proposes a research plan you can edit before it starts, you can interrupt a run, and the report comes with citations or source links.

4. Workflow and automation agents. Zapier’s help center says “Agents are AI-powered assistants that do work on your behalf” (Zapier help). The line between workflow and agent is blurry here: Zapier says it is migrating Agents to “AI by Zapier”, where a migrated agent becomes a Zap with a trigger and a single AI step holding the agent’s prompt, and it has not set a date to turn the standalone product off (Zapier help).

5. Sales agents. Artisan’s home page says its agent Ava “finds leads, enriches them, sends personalized messages, and books meetings” (Artisan). Its pricing page says Ava sends from the customer’s own sending domains in the reps’ names and that calls stay with people: “Ava emails, your reps dial” (Artisan pricing). That is a useful split between what an agent does alone and what a human keeps.

6. Customer support agents. Intercom describes Fin as “the Customer Agent” that handles service, sales, ecommerce and success roles over voice, chat, email, Slack and social (Fin). Its trust page says Fin “escalates to a human agent rather than guessing” (Fin trust). That is a vendor statement; the pages cited give no independent check.

7. General assistants that act for you. OpenAssistant is our example. It works through email, text (iMessage, SMS, WhatsApp), phone calls and Slack, and keeps its own name, email address, phone number, calendar and secure computer (OpenAssistant). In Slack you direct-message it or add it to channels, and it responds only when invoked (OpenAssistant help). It is a general work assistant, not a specialist coding agent, SDR or support-desk product, so a specialist from types 1, 5 or 6 will usually fit those jobs better.

A six-question test: is this product really an agent?

This is our own checklist, built from the shared parts of the definitions above. It is not a standard. Answer yes or no for each product, using the vendor’s documentation.

1. Does it work toward a goal you give it, rather than answer a single prompt? Google Cloud’s definition centers on pursuing goals and completing tasks.
2. Does the model choose the steps and tools? This is the line Anthropic and OpenAI draw between agents and fixed workflows.
3. Does it act in other systems, such as sending, booking, editing or buying, and not only write text? OpenAI’s wording is tasks “on your behalf”.
4. Can it keep working without you watching each step? Claude Code’s cloud sessions and OpenAI’s cloud browser are both documented as running remotely.
5. Does it pause for you at the right moments? Anthropic says agents can “pause for human feedback at checkpoints or when encountering blockers” (Anthropic). Look for approval settings you control per action.
6. Can you see and limit what it can reach? Look for stated permissions, connected accounts you can revoke and admin controls.

How to read the score: count only questions 1 to 4. Four yes answers is an agent in the sense the four definitions share. Two or three is an assistant or a hybrid, which is often what you want for work that touches customers. One or none is a chatbot or a workflow with an AI step. Questions 5 and 6 do not decide whether something is an agent, but they decide whether you should switch it on. Read How to choose an AI agent for the full buying checklist.

Limits of AI agents, and where to read next

Agents make mistakes and can be steered by the content they read. OWASP lists “Excessive Agency” as a risk and names excessive functionality, excessive permissions and excessive autonomy as typical root causes (OWASP). Its entry on prompt injection says it is “unclear if there are fool-proof methods of prevention” (OWASP), and the UK NCSC wrote that such attacks “may never be totally mitigated” (NCSC). Give an agent the smallest set of permissions that gets the job done and keep approvals on for anything hard to undo. Our guide Are AI agents safe? goes through prompt injection, permissions and approval settings.

Vendors also state their own limits. OpenAI’s cloud browser page answers “Can the cloud browser finish every task or transaction?” with “No.” (OpenAI help). For OpenAssistant, the help center says it does not take part in group texts or group WhatsApp chats (OpenAssistant help), does not work in Slack Connect channels with external users (OpenAssistant help), and a Confirmation Window is not the same as Manual approval (OpenAssistant help). Its trust page states a SOC 2 Type II audit and that it never trains on your data (OpenAssistant trust page); those are the company’s own statements.

One more caution on sources: Anthropic’s agent guide carries a note that tooling “has changed since December 2024” (Anthropic), so use it for the concepts, not the tool list. For a ranked view of products by job, see Best AI agents in 2026.

Related reading

• How to choose an AI agent: 10 questions to ask before you buy: Ten questions on job fit, channels, approvals, price, data and exit, with a job-to-product table.
• Are AI agents safe? Prompt injection, permissions and approval settings: Real risks from prompt injection to excess access, the controls to look for and a 10-question checklist.
• Best AI agents in 2026: top picks by job, with prices: Our recommendations for each job (general work, coding, browser use, research, sales, support, enterprise and builders), with vendor prices.
• Best AI agent builders and platforms in 2026: no-code to code: Ten agent builders, including n8n, Copilot Studio, Make, Zapier and CrewAI, with credits, tasks and executions explained.
• Best AI browser and computer-use agents in 2026, compared: Nine browser and computer-use agents compared on documented controls, where the browser runs, approvals and prices.

Frequently asked questions

Is a chatbot an AI agent?

Not by itself. Google Cloud compares agents, assistants and bots by autonomy and puts agents highest (Google Cloud). A product that only replies in a chat window sits at the low end. Once it can act in your other tools toward a goal, the agent label starts to fit.

Is workflow automation the same as an AI agent?

No, according to Anthropic. It defines workflows as paths fixed in advance and agents as systems where the model directs its own process, and calls both agentic systems (Anthropic). Many products mix the two, as Zapier’s move to a single AI step inside a Zap shows.

Can an AI agent act without my approval?

It depends on the product and the settings. OpenAI’s guide says to hand off to a human for high-risk actions such as canceling orders, authorizing large refunds or making payments (OpenAI guide). OpenAssistant lets you set Manual, Confirmation Window or Autonomous per action type (OpenAssistant help).

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

Google Cloud describes an assistant as a type of agent that works with your supervision. In practice the difference is how much it decides alone and how often it checks with you. See the table above for the working meanings we use.

Try OpenAssistant

If you want an assistant that acts for you across email, text, calls and Slack, start with OpenAssistant. The pricing page lists a free Starter plan with 1,000 credits at signup, then 100 bonus credits a day once those are used up, up to 1,000 a month (OpenAssistant pricing). Start with approvals set to Manual and loosen them as you build trust.