A Tour of What Agents Can Do Today

You now know what an agent is: a model running in a loop, using tools, to reach a goal. The next question is the practical one. Which agents can you actually use today, and which one should you reach for when you have real work to hand off?

The answer is less confusing than the product launches make it seem. Almost every agent on the market falls into one of five families: chat assistants that take actions, coding agents, research agents, browser agents, and scheduled or background agents. Each family has a sweet spot and a set of predictable failure modes. This lesson gives you a guided tour of all five, with an honest note on what each is genuinely good at, so that by the end you can match any task to the right kind of agent in a few seconds.

The five families at a glance#

Product names change every few months. The families don't. Here's the map you'll use for the rest of this lesson:

Notice that the families are defined by what the agent can touch, not by which company makes it. The tools an agent has decide what it can do far more than the brand on the box. Many products also span several families: Claude, ChatGPT, and Gemini each offer a chat assistant, a research mode, and more.

Chat assistants that take action#

The chat app you already know has quietly become an agent. When you ask Claude, ChatGPT, or Gemini a question today, it may search the web, read several pages, run a short program in a sandbox (a sealed-off computer where code can run safely), and create a file for you before it replies.

Most chat assistants can now also:

• Read your files: spreadsheets, PDFs, slide decks, images, screenshots.
• Create files: a cleaned-up spreadsheet, a chart, a document, a small web page.
• Connect to your apps through connectors (many built on the Model Context Protocol, or MCP, which you'll meet in Module 4), such as your email, calendar, or cloud drive.
• Remember context through projects and memory features, so you don't repeat yourself every time.

Genuinely good at: quick research questions, drafting and editing, analyzing a single file, explaining things, and thinking through a decision with you.

Where it struggles: work that spans many files, anything that needs to run on your own computer, and long jobs. A chat assistant works inside its own sandbox and the apps you've connected. It can't open the folder of 400 receipts on your laptop.

Look at the attached spreadsheet of our Q3 expenses. Find the five
largest categories, flag any line items that look like duplicates, and
make a bar chart of spend by month. Then write three sentences I could
paste into an email to my manager summarizing what changed vs. Q2.

Coding agents: not just for code#

Coding agents like Claude Code, OpenAI Codex, Cursor, and GitHub Copilot were built to write software, and they're very good at it: they read your project, edit files, run tests, see what broke, and fix it. Module 2 is devoted to using them well.

Here's what surprises most people: a coding agent is the most capable general-purpose agent you can run on your own computer. Its tools are reading files, writing files, and running commands, and those three tools cover an enormous amount of non-programming work. Renaming and sorting 300 photos by date. Merging twelve CSV exports into one clean spreadsheet. Turning a folder of meeting notes into a summary of decisions. Converting a pile of Word documents to Markdown. You'll do exactly this kind of task in Your First Agent Task.

They come in three shapes, which you'll explore in The Coding Agent Landscape:

• Terminal agents run in a command-line window inside a folder you choose.
• IDE agents live inside a code editor, where you can watch every change appear.
• Cloud agents work on a copy of your project on a remote machine and come back with the finished changes for you to review.

Genuinely good at: anything where the work lives in files and success can be checked, by running tests, re-reading output, or opening the result. That "can be checked" part is why coding agents became the first agents people trusted with real work.

Where they struggle: tasks that need judgment they can't see (your taste, your client's preferences), and anything outside the files and tools you've given them. They also act on your real files, so the habits you'll learn about git and permissions matter.

Research agents: many sources, one report#

Ask a normal chat assistant a hard question and it might run three searches. Turn on a research mode (often called "deep research") and the agent behaves differently: it plans a set of questions, runs dozens of searches, reads many pages, follows leads it discovers along the way, and writes a structured report with citations. A run might take five minutes or half an hour.

Genuinely good at: surveying a topic you don't know, comparing options (vendors, tools, approaches), gathering scattered facts into one place, and finding sources you wouldn't have found yourself.

Where it struggles: research agents can only read what's publicly reachable, so paywalled journals and private data are out of view unless you connect them. More importantly, they make citation mistakes: a claim may cite a real page that doesn't actually say what the report claims. They also tend to weigh a popular blog post the same as a primary source unless you tell them otherwise.

Browser agents: clicking so you don't have to#

A browser agent operates websites the way you do. It looks at the page (literally, through screenshots, or by reading the page's structure), then clicks buttons, fills in forms, scrolls, and moves between tabs. Claude in Chrome, ChatGPT's agent mode, and Gemini in Chrome are examples. A close cousin, the computer-use agent, does the same thing for a whole desktop: any app, not just a browser. Module 7 covers both, starting with Browser Agents.

Genuinely good at: chores on websites that offer no better way in. Filling out a long form with information from a document. Collecting prices from five retailers that have no API (a programmatic interface for software to talk to the site directly). Working through a clunky internal portal.

Where it struggles: this is the newest and least reliable family. Browser agents are slow compared with a person who knows the site, they get confused by pop-ups, logins, and unusual layouts, and they sometimes click the wrong thing. They also carry a special risk: a web page can contain hidden text written to trick the agent into doing something you didn't ask for, an attack called prompt injection that Module 10 covers in depth.

The rule of thumb: if a better path exists, use it. If a service has an official connector, an API, or a data export, an agent using that path will be faster and far more reliable than one clicking through the website.

Scheduled and background agents: work while you're away#

Every agent above can run while you watch. The last family runs while you don't.

• Scheduled tasks: several chat assistants let you save a prompt that runs on a schedule, such as a Monday-morning news briefing or a daily check of a competitor's pricing page.
• Cloud and background agents: you hand a coding agent a task from a web page, an issue tracker, or chat, and it works on a remote machine for minutes or hours, then hands you the result to review. You can start three tasks before lunch and review three results after.
• Agents on a timer: coding agents can run non-interactively from a script, so you can schedule them like any other program. Module 8 walks through this in Scheduled and Background Agents.

Genuinely good at: recurring work with a stable shape, and long tasks where you'd rather review a finished result than supervise each step.

Where they struggle: nobody is there to answer questions or catch a misunderstanding, so a vague instruction produces a wrong result every day until you notice. Background agents need tighter instructions, narrower permissions, and a clear way to tell you what happened.

Matching tasks to agents#

When you have a task in mind, ask two questions: where does the work live, and how much supervision do you want to give? The table below covers most cases.

Two warnings apply across every family. First, every agent sounds equally confident whether it did the job perfectly or quietly skipped half of it, so you check the result, not the summary. Second, the families are converging. Chat apps keep gaining coding and browsing abilities, and coding agents keep gaining research tools. Don't worry about the "correct" product; ask whether the agent can touch what the task needs and whether you can check its work.