AI agents, explained
Short, plain answers to the questions people ask before they hand work to an AI agent.
What is an AI agent?An AI agent is software that uses an AI model to carry out a task step by step with tools, instead of only answering. Examples, limits and how it differs from a chatbot.- AI agent vs chatbot: what's the difference?A chatbot answers in a text box. An AI agent has tools, a place to work, memory and a schedule, so it does the task itself. A side-by-side table.
- How do AI agents work?An AI agent runs a loop: the model picks a step, a tool carries it out, the model reads the result and picks the next step, until the task is done.
- What is an AI teammate?An AI teammate is an AI agent that stays: it has a role, remembers your work, starts tasks without being asked and works alongside people.
- Why give an AI agent its own computer?An AI agent with its own computer works apart from your devices, keeps going while you're away, keeps its files and sign-ins, and can be rolled back.
- Is it safe to give an AI agent access to my email and accounts?It can be, if the agent asks before sending, can't see your passwords, treats message content as untrusted, and you can see and undo what it did.
- Which AI agents ask before they act?Most AI agents ask before some actions. They differ in which actions, whether you can set the rules, whether payments always ask, and whether you can undo.
- Self-hosted or hosted AI agent: which should you choose?Host it yourself for control over data, models and cost. Use a hosted agent for no setup and no upkeep. A table of trade-offs and a way to decide.
How do I self-host an AI agent?To self-host an AI agent you need a machine with Docker, an AI model key or a local model, and about 15 minutes. The steps, with Casama as the example.- Can AI agents run on my own models?Yes. An AI agent can use a local model through Ollama, LM Studio or vLLM if the agent software accepts an OpenAI-style endpoint. What works and what doesn't.
- How do I run AI agents on my own server from my phone?Run the agent software on a computer at home, give it an address your phone can reach without port forwarding, and sign in with a native app.
- What is MCP, and what is an MCP client?MCP (Model Context Protocol) is a standard way for AI apps to use outside tools and data. A server offers tools; a client is the AI app that uses them.
- What is an AI employee?An AI employee is an AI agent given a standing job, with tools, memory, a schedule and limits. What AI workers and AI interns can do, what they cost, and where they fail.
- Are there free AI agents?Yes. Open-source AI agents are free to run and some hosted ones have free plans, but the AI model usually costs money unless you run one locally.
- Which AI agents have iPhone and Android apps?Which AI agents have real phone apps, which are in early access, and which you reach through chat apps instead. Seventeen agents checked, with sources.
- Which AI agents let you choose the model or use your own API keys?Open-source AI agents usually take your own API keys, subscriptions and local models; hosted ones usually pick for you. Seventeen agents checked, with sources.
- Which AI agents can run tasks on a schedule?Most AI agents can run on a schedule now. They differ in what else can start them, the limits, whether your computer must stay on, and what still asks first.
- Which AI agents don't train on your data?Which AI agents say they don't train on your data, which train by default until you opt out, and which don't say. Seventeen agents checked, with sources.
- How much does an AI agent cost a month?From $0 for free software and a local model, to $15 to $200 a month for most hosted AI agents. What you pay for, and the published prices of 17 agents.
- What is a computer use agent, and which ones let you take over?A computer use agent works a computer the way you do: screen, mouse, keyboard, browser. Where it runs, why it matters, and which agents let you watch and take over.
- Is OpenClaw safe, and is it free?OpenClaw is free and MIT-licensed. It can be run safely, but its sandbox and approvals are off by default. What that means, and how to set it up more safely.