AI Agents.

What an agent actually is

"Agent" is the word everyone is using right now, and most people I talk to are not quite sure what it means. The single most useful definition is also the shortest. A chatbot writes back. An agent goes and does the thing.

Imagine you ask the same question two different ways. First as a chat: "Can you find three dentists near my zip code who take my insurance and have weekend hours?" The chatbot writes back a polite paragraph explaining that it can't actually browse the web for you, suggesting search terms, and wishing you luck. You go open Google.

Now imagine you ask an agent the same thing. It actually opens a browser, runs the searches, visits the dentist websites, reads the hours and insurance pages, and twenty minutes later puts a small table in front of you: three names, three phone numbers, three insurance confirmations, the source link next to each row. You didn't open Google. You read a table.

That's the whole shift. Agents don't just compose words. They take actions in the world on your behalf. They click, they fill in forms, they read pages, they send messages, they schedule, they check back tomorrow. They use the same underlying AI brain a chatbot does, plus a set of tools that let that brain do things instead of just talk about doing them.

This distinction matters because the questions you ask the two are different. With a chatbot, you're shaping a draft together — write me, summarize this, explain that. With an agent, you're delegating a task and getting back a result. "Watch this site and tell me when the price drops below $40." "Every Monday at 7am, summarize what happened in our family group chat last week and put it in my calendar as a note." "Find me three quotes for a new dishwasher with the model I want, and email me the table." Chatbot prompts are usually conversations. Agent prompts are usually instructions.

The four flavors you'll meet

The word "agent" gets used for at least four different shapes of thing, and that's part of why the term feels slippery. Once you can recognize them by name, half the confusion goes away. You may already be using two of these without realizing they qualify.

Most "AI agents" you hear about are some mix of these four. Claude Cowork is a task agent that can use connectors and a browser. ChatGPT Tasks is a task agent. ChatGPT Agent is a browser-use agent. Microsoft 365 Copilot is an assistant-in-an-app plus a connector agent. You don't need to memorize this map, but you'll find that almost every new "agent" announcement slots neatly into one of the four buckets.

The three questions before you trust one

Before you let any agent do anything for you, get into the habit of asking three plain questions. They're the same three questions you'd ask before handing your house keys to a contractor, and the answers determine how much rope you should give the thing.

1. What can it see? Just the message I'm typing? Or my whole Gmail? Or the file I'm working on? The browser tab I have open?
2. What can it do on my behalf? Just talk? Click around the web? Send messages? Spend money? Change settings in my account?
3. Where does it stop on its own? Does it ask before sending an email, or does it just send? Does it check in before spending money, or click "buy"? When it's done, does it stop, or does it keep going until I tell it to stop?

Every well-built agent answers these clearly somewhere in its setup. If you can't find the answers in plain language, that's a signal to start small. Most of the bad agent stories I've heard come from people who skipped one of the three questions because the answer wasn't obvious and they didn't push.

If you want a deeper read on the safety side of all this — what gets logged, what gets used to train models, what's actually private when you use a connector — the AI Safety guide is the practical version of that conversation. It's worth reading once before you wire any agent into your email or calendar.

The weekly brief project

Here's the project. Pick one topic you genuinely care about and would like to keep up with, but never quite do. A hobby you're rebuilding. An industry you're tracking for work. A health condition someone in your family has. A team or a band or a stock or a kid's school subject. Anything where you've thought "I should pay more attention to this" and then haven't.

You're going to set up an agent that, every Monday morning at 7am, sends you a short brief on what happened in that topic over the past week. Three or four bullets. One link per bullet. Long enough to feel useful, short enough to read while the coffee's brewing.

Three different platforms can do this today, and the shape of the prompt is almost identical in all three. Pick whichever one you already have access to, and skip the others.

Variant A — ChatGPT Tasks

Available to anyone on a paid ChatGPT plan. You set up a "task" that runs on a schedule and emails you the result.

Variant B — Claude Cowork

Available with a Claude paid plan. Cowork can run scheduled tasks with browsing turned on, which means the brief will reflect actual recent web sources rather than only what the model already knows.

Variant C — Gemini scheduled prompts

Available with a Google account using the Gemini app or Workspace. The mechanics are slightly different — Gemini calls these "Scheduled actions" — but the prompt is the same.

What week one will feel like

The first brief that lands will probably be good but not great. The topic will be a little too broad or too narrow. The "what's new this week" framing will surface things that aren't actually news. The links might point to a roundup post you don't trust. This is normal and is the whole point of week one. You're calibrating.

By week two or three you'll have made three or four small adjustments — narrowing the topic, asking it to skip certain sources, requesting a different format — and the brief will start landing in your inbox in a shape you actually want to read. That's when the project earns its keep. You'll have a thing arriving every Monday morning, with no effort from you, that keeps you a little more current on something you cared about but didn't have time for. That feeling — useful work happening while you were doing something else — is what makes agents agents.

A browser-use errand

The weekly brief shows the "agents that run on a schedule" shape. The other shape worth meeting is the "agent that opens a browser and goes and looks at things for you, right now." This one is more dramatic to watch, slower to run, and the right tool for a narrower set of jobs.

The right job for browser-use is something that would take you fifteen to thirty minutes of clicking through similar pages and copy-pasting bits into a doc. Comparison shopping. Researching three options that all look similar. Filling out a multi-step form where each step takes you to a new page. Pulling structured information out of an unstructured site.

The wrong job for browser-use is anything that involves your bank account, your investment account, your government accounts, or anything where a wrong click would be expensive or irreversible. Stay on the safe side of the line. The agents are getting better at this; the time to trust them with money will come; today is not that day.

What you'll get back is a table. Some cells will be confidently filled in. Some will say "couldn't determine" because the site hid the price behind a sign-in. Some will be a few dollars off because the site has dynamic pricing tied to your zip. All of this is normal. The win is that you have a real comparison table in front of you that took twelve minutes to produce, during which you did something else. The next time you'd have done this task, you'd have spent thirty minutes and gotten a worse table.

Once you've run one of these errands successfully, you'll start spotting candidates in your week. Comparing four health insurance plans your employer offered. Finding the cheapest fare across three travel sites. Pulling the contact phone numbers off the websites of every preschool within five miles. Looking up the warranty terms on a product across five retailers. Anything where the underlying work is "open ten similar tabs and pull the same three pieces of information out of each."

Patterns that keep working

If you only remember a handful of things from this whole guide, remember these. They're the habits that pay off across every flavor of agent, every platform, and every topic.

1. Small, repeatable, reviewable. The agents that earn their keep are the ones doing a small specific job over and over, where you read the output. Big vague tasks fail loudly. Small specific tasks succeed quietly. Start small.
2. Give it a format, not just a topic. "Tell me about solar batteries" produces a wall of text. "One headline sentence, three bullets each with one link, one watch-next-week line" produces something you can read in ninety seconds. The format is most of the prompt.
3. Tell it what to do when there's nothing to say. The single most common failure mode of scheduled agents is padding. They will manufacture interestingness on slow weeks if you don't give them permission to admit a quiet week. "Say so honestly in one sentence" prevents this.
4. Drafts, not sends. Notes, not bookings. Until you've watched an agent do a class of task five or six times, ask it to draft the email and leave it in your drafts, not to send it. Ask it to put the booking details in a doc, not to book the trip. The work is the same. The blast radius is not.
5. Edit the instructions at the source. When a scheduled agent's output drifts, don't try to correct it in the chat. Open the task, edit the prompt, save. The next run will be better. This is the agent equivalent of "edit and resend."
6. Verify load-bearing links and numbers. A brief in your inbox is useful. A brief in your inbox you'd act on without checking is a different category. If you're going to repeat a fact, click the source first. This is the cost of using a tool that's fluent and occasionally wrong; you can pay it in two minutes a week.

Patterns that fail

The flip side. These are the recurring ways agents disappoint people, and they're avoidable once you've named them.

• The big vague task. "Plan my retirement." "Organize my email." "Manage my calendar." Agents are bad at problems with no edges. Find the smallest concrete slice of the big thing — "draft three retirement scenarios into a doc using the numbers from this PDF," "label every email from my kids' school with this tag" — and run that. The full thing rarely works; a thin slice often does.
• Trusting the first run. First runs are calibration runs. The first weekly brief, the first errand, the first scheduled task — read those with the skepticism of a job interview, not the warmth of a delivered product. By the third run you'll know whether to trust it. Until then, verify.
• Connecting everything at once. The temptation when you first wire up a connector agent is to give it access to Gmail, Calendar, Drive, Slack, and the kitchen sink, then ask it a big question. Don't. Add one connector at a time, give it a small task scoped to just that connector, watch it work, then add the next one. You're not being slow; you're building a mental model of what the agent does well in each context.
• Letting it act on your behalf without a review step. Sending emails. Posting messages. Booking things. Spending money. There are agents that will do all four, and there are tasks where letting them is the right call — eventually. Not on the first ten runs of any task. The "draft, don't send" version of a workflow is almost always one click further from disaster than the "send" version, at no real cost to you.
• Asking the same agent for a second opinion on its own work. If a brief feels off, don't ask the same agent "are you sure?" — it has the same institutional bias toward agreeing with itself that a chatbot does. For a real second opinion, run the same prompt in a different tool, or open a fresh task and look at the disagreement.
• Mega-prompts where a few iterations would do. Trying to specify every constraint up front in one enormous task description usually goes worse than running the task, reading what comes back, and refining the prompt over two or three weeks. The agent is going to run repeatedly anyway. Use that.

Where to go next

If the weekly brief is landing in your inbox and you've run an errand or two, you've already crossed into the part of agent use most people never reach. From here, the next step depends on which platform you settled on.

• If you picked Claude Cowork. The Cowork guide goes deep on connectors, computer-use sessions, dispatch (running multiple tasks in parallel), and the enterprise controls that gate them. That's where you'll learn how to wire Cowork into Gmail or Calendar safely, and how to scale from one weekly task to a small portfolio of them.
• If you picked ChatGPT. The ChatGPT guide covers Tasks, Agent (browser-use), Projects, custom GPTs, and the features that matter once you've outgrown the basic chat. The boundary between "ChatGPT" and "ChatGPT Agent" is where most people lose the thread; the guide walks that line.
• If you picked Gemini. The Gemini guide covers the app, the Workspace integrations, scheduled actions, and the Android assistant role. Gemini's biggest superpower is the Workspace connector story — your Gmail, your Calendar, your Drive — and the guide walks through wiring those up.
• If you're still on the fence. The Everyday AI guide is the chatbot-shaped version of "useful AI in real life." Going deeper there before agents is a perfectly reasonable order. Chatbot fluency makes agent prompts much better, because every agent prompt is really a chatbot prompt with extra rules about what to do at the end.
• Always worth the read. The AI Safety guide. If you've started connecting agents to your real data — email, calendar, drive — read this before you connect the next one.

Agents work best when you think of them as a junior teammate who is fast, tireless, fluent, occasionally wrong, and has exactly the permissions you gave them and no more. Treat them that way and they earn their keep. Treat them as oracles and they'll let you down. — TWD

Set up the weekly brief this Sunday. Pick a topic that matters to you. Run it for three weeks before you decide if it's working. The skill is small and stackable, and the only real way to learn it is to have one of these running in your life.