An agent is a tool that clicks, not a chat that promises
AI agents are systems that take a goal and then use tools: a browser, a shell, a calendar, a file, another model. The pitch is that they do the thing, not merely describe it. The reality in 2026 is narrower. A useful agent does one chore with a visible trace and a stop button. A bad agent does ten chores you did not confirm and writes a confident log.
This is a way to tell those apart and to pick a tool that actually finishes a job you have. It is not a leaderboard of startups. Products appear and vanish. The tool directory is updated more slowly than a launch thread, on purpose. The glossary is where "agent," "tool use," and "workflow" get plain definitions. Read those before you buy a noun.
What "actually does things" should mean
A agent earns the name when you can point to a side effect outside the chat: a file changed, a draft sitting in your outbox unsent, a spreadsheet cell filled, a ticket opened. If the side effect is only a paragraph, you have a chatbot. Use a chatbot. They are simpler and they fail more honestly.
The side effect needs a boundary. "Organize my files" is not a boundary. "Rename these ten screenshots using the date in the filename, and put the plan in a text file I approve" is a boundary. Agents are good at the second and expensive at the first, because the first includes your tax folder.
Classes of tools that exist
Editor and coding agents edit a repository and run tests. Judge them by the diff. The coding assistant guide is that class. Do not confuse it with a browser agent.
Browser agents click through websites. They are useful for a repetitive form you are allowed to automate and dangerous on anything with a payment or a delete button. Watch the first run. If you cannot watch, do not run it.
Inbox and calendar agents draft or schedule. The email guide keeps the send button in your hands. A calendar agent that emails other people needs the same restraint.
Research agents open many pages and file a memo. Treat the memo as unsorted links until you click them. The Perplexity practice is the non-agent version of the same discipline. An agent that browses more is not an agent that checks more.
Computer-use demos that move your mouse are impressive and brittle. Give them a spare account, not your primary login, and not an account that can pay.
A selection table
| You want this side effect | Hire this class | Refuse this pitch |
|---|---|---|
| A reviewed code change | Coding agent with a diff | "It shipped to production" |
| A folder of renamed files | A script, or a agent on a copy of the folder | Access to your whole home directory |
| A packet of links | Research assistant with citations you open | A finished market report |
| A draft email | Suggested reply | Auto-send to customers |
| A weekly metric pulled from a export | A real query or a small script | A agent that screen-scrapes your bank |
The script row is intentional. Many agent demos are a shell script with a conversation in front. If you can write the script, or a teammate can, you get a thing you can diff in git. Prefer that for anything you will run twice.
How to trial an agent in an afternoon
- Write the goal so a stranger could tell whether it finished.
- Give it a copy of the data, or a test account, never the only copy.
- Watch the trace. If there is no trace, stop.
- Approve each new kind of action the first time: send, delete, pay, invite, deploy. Default those to off.
- When it finishes, check the side effect yourself before you thank it.
If it fails, resist the urge to widen its permissions. Widen the permissions and you will get a larger failure. Narrow the goal instead.
Stop conditions belong in the setup: maximum steps, maximum minutes, no spending. A runaway browser session is a cost even when it does not succeed. The subscription guide includes metered agent runs. They are easy to leave looping.
What agents are bad at
They are bad at goals that require taste halfway through, because they will not stop to ask unless you forced a checkpoint. Put the checkpoint in. "Show me the list before you delete" is a professional requirement, not a lack of faith.
They are bad at logging into sites that do not want bots, and you should not help them evade that. If a task requires breaking a terms page or a captcha puzzle meant to stop automation, it is not a personal productivity project.
They are bad at being your only memory. When the run ends, copy the result into the system you already trust. A agent transcript is not a filing cabinet.
A reasonable 2026 stack
One chatbot for drafts. One coding assistant if you write code. One browser or file agent on a test folder, used rarely, with a stop button. That is a lot of power. A fourth "life operating system" that holds your mail, bank, and files is a demo you do not need this year.
Revisit the choice when a chore is so repetitive that the trace is boring. Boring and correct is the adoption signal. Exciting and almost correct is the signal to keep watching.
Frequently asked questions
Q: What are the best AI agents that actually do things?
A: The ones whose side effect you can check: a diff, a renamed copy of a folder, a draft that does not send. Pick the class that matches the side effect. Ignore a ranking that scores demos.
Q: Are agents better than chatbots?
A: They are better when you need a click or a file change and you can supervise it. They are worse when you needed a paragraph, because the extra permissions add ways to fail.
Q: Can I let an agent use my computer overnight?
A: Not with your primary accounts, and not with pay or delete enabled. Overnight runs belong on a copy of the data, with a step limit, if they belong anywhere.
Q: Do I need several agents?
A: You need the class that matches your chore. One is plenty until that one is boring and correct.
Q: How is an agent different from a macro?
A: A macro does exactly what you recorded. A agent improvises toward a goal. Improvisation is why you watch. If the task is stable, a macro or a script is safer.
Q: What permissions should stay off?
A: Pay, delete, send to other people, deploy, and access to password stores. Turn one on only for a goal that cannot be done otherwise, on a test account, once.
Q: Why did it stop in the middle?
A: It hit a login, a changed button, or a ambiguity you would also have paused on. That pause is a success if it asked you. If it guessed, narrow the goal.
Q: Will agents replace the tools in the directory?
A: They sit on top of tools and browsers. The underlying mail client, editor, and file still matter. Browse the directory for those. Use a agent only when the glue itself is the chore.
Demand a side effect you can see
Give one agent one bounded job this month, on a copy of the data, with send and pay turned off. If the trace is clear and the result checks out, you have found a tool that does a thing. If the pitch needs your whole computer and a weekend of faith, keep the chatbot. The blog covers the narrower automations, email and code and research, which solve most of the jobs an agent demo is filmed doing.
