AI Agents Examples: 20 Agents by Type, With Sources (2026)

Below, each example links to the official product page or documentation we checked on 2026-10-25, and every price comes from the vendor's pricing page on that date. To be clear about method: this is a sourced overview, not a hands-on test. We did not run these agents on real tasks for this article, so we describe what each vendor documents and point out where those claims need your own trial.

What counts as an AI agent (and what does not)

Many lists of AI agents mix three different things: chatbots that only answer, automations that follow fixed steps, and agents that decide. Separating them saves you from paying agent prices for a job a simple rule could do.

Anthropic's engineering guide, Building effective agents, draws the line clearly. Workflows are "systems where LLMs and tools are orchestrated through predefined code paths." Agents are "systems where LLMs dynamically direct their own processes and tool usage." The n8n documentation says the same thing in plainer words: a chain follows a predetermined sequence, while an agent "uses a language model to determine which actions to take."

So a practical test for any product calling itself an agent:

• Does it choose its own next step? If every step is drawn in a flowchart by you, it is a workflow, even if one step calls an AI model.
• Can it use tools? A browser, a terminal, your inbox, an API. Without tools, it can only talk.
• Does it check results and try again? Agents loop: plan, act, look at the outcome, adjust.

Our view: "agent" has become a marketing label. Plenty of products sold as agents are well-made workflows, and that is not a criticism. Fixed steps are cheaper, faster and easier to audit. Anthropic's own guide recommends "finding the simplest solution possible, and only increasing complexity when needed," and notes that agentic systems "often trade latency and cost for better task performance."

The five classic types of AI agents, with examples

Before LLMs, computer science already had a taxonomy of agents. It still helps, because modern products map onto it. IBM's summary of AI agent types lists five, each with textbook examples.

The best-known real-world utility-based agent is probably a driverless car. Waymo describes itself as "the world's first autonomous ride-hailing service" and, on its homepage on 2026-10-25, listed 15 metro areas currently serving riders (including Phoenix, Los Angeles, the San Francisco Bay Area, Austin and Miami) and "20M+ rides served." That is an agent in the strict sense: it perceives, plans and acts in the physical world with no driver.

Most software agents you can buy today are goal-based. You state the goal, the model plans, and tools do the work. The "learning" part is usually weaker than marketing suggests. Many products remember your preferences, but few genuinely improve their strategy from outcomes without a human reviewing the results.

Everyday examples of AI agents for email, meetings and the web

This is the group most people mean when they search for AI agents examples: assistants that take routine work off your plate. For a ranked shortlist rather than examples, see our best AI agents page.

• Lindy calls itself "the AI employee that plugs into your tools, learns how you work, and takes real work off your plate." Documented tasks include drafting and replying to email, sharing meeting notes after calls and scheduling. Its pricing page lists 1,500+ integrations, MCP support and built-in approvals. New accounts get 50 USD in free credits that last 7 days, and the Team plan costs 49.99 USD per month for 5,000 credits, according to its pricing page.
• Zapier Agents let you "equip your Agents with company knowledge and have them do work across 9,000+ apps." Example agents on the page include a Meeting Prep Agent that builds a dossier on attendees, a Lead Qualification Agent and a Zendesk Ticket Agent that answers common questions and escalates complex ones. Zapier's pricing page shows a free plan with 100 tasks per month and Professional from 29.99 USD per month (19.99 USD per month billed yearly).
• Comet is Perplexity's browser, pitched as "the browser that works for you," available for Mac, Windows, iOS and Android. Browser agents like this act on the pages you visit, which is powerful and also the riskiest place for an agent to work. Our Perplexity profile covers the rest of its product line.

A dated lesson from ChatGPT

ChatGPT is the clearest proof that agent products change fast. OpenAI's help article on ChatGPT agent, checked on 2026-10-25, opens with: "ChatGPT agent is no longer available. Use ChatGPT Work for longer, multi-step tasks and finished deliverables." The same page still documents the safeguards OpenAI built in, including "user confirmations for high-impact actions," prompt injection monitoring and a "watch mode" requiring supervision on certain sites. See the OpenAI Help Center article.

If you are still choosing a general assistant before adding agents on top, our ChatGPT vs Gemini comparison and the ChatGPT alternatives page cover the base models.

Coding agents: the most mature category

Coding is where agents work best today, for a structural reason Anthropic's guide points out: code can be checked automatically. Tests either pass or fail, which gives the agent a clear feedback signal. Its guide describes coding agents resolving GitHub issues that need edits across many files, while noting that human review remains important.

The workflows differ more than the table suggests. Jules shows the classic agent loop on its homepage: it clones your repository to a cloud VM, drafts a plan you approve, shows a diff of the changes, and then you approve the PR and publish it on GitHub. Claude Code's page says it "plans, writes code, runs tests, and opens pull requests," and that it "asks for permission before making changes to your files or running commands." Devin Desktop, which windsurf.com now redirects to, is another option: its pricing page lists a Free plan with a "light quota to code with agents" and Pro at 20 USD per month.

Our take: the gap between these tools is less about raw model quality and more about where they fit your habits. Terminal-first developers lean toward Claude Code (see our Claude profile) or the Codex CLI. Teams that live in GitHub issues may prefer an agent you can assign by label. Compare current model rankings on our AI leaderboard, and see the full category in AI for coding.

Business workflow agents you build yourself

The fourth group is less a product than a pattern: you assemble an agent from a builder, connect your own apps and decide exactly which actions need a human. Common business examples:

• Customer support triage. Read incoming tickets, answer the routine ones from your docs, escalate the rest with a summary. Anthropic's guide calls support a natural fit because success is measurable (was the issue resolved?).
• Lead qualification. Enrich a form submission, score it and alert sales only for good fits.
• Meeting prep and follow-up. Research attendees before a call, then turn notes into tasks. Pair with tools from our AI note-taking list.
• Back-office classification. Sort expenses, route invoices, tag documents.

Notice what is not in that table: Microsoft Power Automate. Its desktop flows are classic RPA (robotic process automation, meaning software that clicks through apps by recorded rules). In IBM's terms that sits closest to a simple reflex agent, and for stable, repetitive screens fixed rules are usually cheaper and easier to audit than an LLM agent.

For a broader shortlist, see AI for automation, or compare cheaper options on our Zapier alternatives page.

How to choose (and when not to use an agent)

After reading through the documentation for these 20-odd examples, our selection advice comes down to four questions.

1. Can you write the steps down? If yes, build a workflow. It is cheaper, faster and repeatable. Use an agent only for the step where judgment is needed.
2. Can the result be checked automatically? Code with tests, a form that validates, a ticket marked resolved. Agents do best where success is measurable.
3. What is the worst action it could take? Sending email, paying, deleting and posting should require approval. OpenAI built confirmations for high-impact actions for a reason, and Lindy, n8n and Claude Code all advertise approval or permission controls.
4. Who pays per run? Credits, tasks and executions are billed differently. Run a small pilot and read the usage log before scaling.

The contrarian point we keep coming back to: the best first agent for most teams is narrow and boring. One inbox, one ticket queue, one repository. Broad "do anything" agents make impressive demos, but they are harder to supervise, and every new tool you connect is another surface for prompt injection.

If you are new to AI tools in general, start with our guide on how to use AI, then browse free AI tools to experiment before paying. Developers who want to build their own agent from scratch can read our AI API guide.

Keeping this list current

Agent products rename, merge and retire features faster than almost any other software category, as the ChatGPT example above shows. We keep tool profiles in our AI tools directory with a source link and a "checked on" date beside every price, and facts older than 90 days are flagged for re-check. For model scores, and for the Bahasa Indonesia benchmark (NS-ID) we are building, see our benchmarks pages.

When you pick an agent from these examples, run it on one real task you already know the answer to. That single test will tell you more than any list, including this one.