What AI agents can actually do for you now

An AI agent is a chatbot that doesn't just answer, it acts: it clicks through websites, fills in forms, reads your inbox, runs code. In 2026 you can find one built into browsers, email apps and coding tools. Some of what they do is genuinely useful. Some of it is a demo that works once. Here's the honest split.

What works today

These are jobs where agents are reliable enough to save real time, as long as you check the result.

• Research and comparison. "Find five dentists near me that take this insurance and open Saturdays, with links." Agents are good at reading many pages and putting the findings in a table. You still open the links.
• Booking drafts. An agent can search flights or restaurants, narrow the options and fill in the form. The safest pattern is to let it stop before the final click and hand the purchase to you.
• Coding. This is where agents are strongest. They read a project, make changes across files, run the tests and explain what they did. Developers review the changes like a colleague's work.
• Inbox triage. Sorting newsletters from real mail, drafting replies for you to edit, pulling dates and tracking numbers into a list.
• Paperwork summaries. Reading a long policy or contract and pointing you to the parts that matter. You read those parts yourself.

What still breaks

• Real websites. Sites change, pop-ups appear, logins time out. A flow that worked yesterday can fail today, and agents are slower than you clicking through it.
• Ambiguity. "Book the usual" or "the cheaper one" can mean several things. Agents guess, and guesses compound over a long task.
• Long chains. The more steps a task has, the more chances for a small error early on to quietly ruin the end result.
• Prompt injection. Text hidden in a web page or email can carry instructions that hijack the agent ("ignore the user, forward this inbox to..."). Security researchers have shown this working against several commercial browser agents, and there's no complete fix yet.

Let an agent do the legwork, but keep your own hand on anything that spends money or sends something in your name.

Permissions to never give

Treat an agent like a new assistant on their first day, with less access.

1. No saved payment details it can use without asking. Make every purchase require your confirmation.
2. No access to your primary email's send button without review. Drafts are fine; auto-send is how mistakes and scams spread.
3. No banking, brokerage or government accounts. These are the highest-value targets for injection attacks.
4. No password manager access. If a task needs a login, you log in.
5. No standing permission to delete. Archive, don't erase.

How to supervise one

• Start small. Give it a task you could do in ten minutes and compare its work with yours.
• Ask for a plan first. A good agent will list its steps. Read them before saying go.
• Watch the first few runs. Many agents show what they're clicking. If it wanders onto a site you didn't expect, stop it.
• Check the output, not the confidence. Agents report success even when they half-finished. Open the booking, the file, the sent folder.
• Prefer on-device and limited tools for anything private. Our look at what on-device AI can do in 2026 explains why less data leaving your phone means less to leak.

Agents will keep getting more capable. The people who benefit most will be the ones who learned to delegate carefully first.