What is an AI agent? A plain-English guide with real examples

A chatbot answers. An agent does the work. That one difference is behind most of the excitement around AI agents, and also most of the confusion. Here is what an AI agent actually is, what it can do today, where it still struggles, and how to start using one.

Chatbot vs agent

When you ask a chatbot for the top story on Hacker News, it tells you what it remembers, which may be months old. When you ask an agent, it opens the site, reads the front page and tells you what is there right now.

An agent has four things a chatbot does not:

1. A job. A standing role, like "morning briefings" or "keep an eye on the team's issues", with instructions it follows every time.
2. Tools. Access to your apps, such as Gmail, Calendar, Slack, Notion, GitHub and Linear, so it can look things up where your work lives.
3. A computer. A real web browser in the cloud, so it can use websites that have no app connection: search, click through, fill in forms and read the result.
4. A schedule. It can do the job every weekday at 9, or every Friday afternoon, and report back without being asked.

Real examples

These are the kinds of jobs people give agents in OneSpace:

• A morning briefing. "Every weekday at 9, open Hacker News and tell me the top AI story with its points and comments."
• An inbox sweep. "Each Monday, go through my inbox and list what still needs an answer."
• Tomorrow, today. "Check my calendar every evening and tell me what tomorrow needs."
• A price check. "Find the cheapest 6 ft USB-C cable on Amazon and show me before you buy."
• A team update. "Every Friday, sum up what the team shipped in Linear and GitHub."

You can also put several agents in a room and let them hand work to each other: one researches, another writes, a third reviews.

Staying in control

An agent that can act needs limits. Here is how OneSpace handles the three that matter most:

• Irreversible steps wait for you. Before an agent places an order, sends a message or does anything else that cannot be undone, it stops and asks.
• Sign-ins stay with you. When a website needs you to sign in, the agent hands the screen to you. It never asks for your password.
• You can watch and take over. You can see the agent's screen as it works, take control at any moment, and hand it back when you are done.

Scheduled routines are read-only by default: they look things up and report back, and do not send, post or change anything unless you allow it.

Where agents still struggle

Agents are useful today, but they are not magic, and it helps to know the limits:

• Websites change. A redesigned page or a pop-up can slow an agent down. A good agent tells you what it found rather than pretending it succeeded.
• Some sites block automation. Captchas and bot checks are built to stop exactly this, and an agent should not try to get around them.
• Vague jobs get vague results. "Keep me updated on AI" is hard to do well. "Tell me the top AI story on Hacker News with its points" is easy. Be specific about what to check and what to report.

How to start

Pick one small, repetitive thing you check by hand every week. Write it down as a job with a clear output ("tell me X, with Y"). That is your first agent.

In OneSpace, agents are on the Pro plan (up to 10 agents) and Max (up to 25). Each gets a name, a job, instructions, and the model you choose for it. See how agents work, the apps they can use, and the plans.