From answering to acting
The agent loop
An agent works in a loop: read the goal, plan a step, use a tool (search the web, open a file, click a button, send a message), look at what happened, then decide the next step. It keeps going until it thinks the job is done or it gets stuck. A chatbot stops after one reply; an agent may take fifty actions.
Where you meet them
By late 2026 agents are built into mainstream products: ChatGPT Work, Claude Cowork and Claude in Chrome, Microsoft's Copilot agents, Google's Gemini agent features, and open-source agents such as Hermes Agent that you run yourself. Many can work in the background or on a schedule while you do something else.
What they're good at
Repetitive, multi-step computer work with a clear finish line: gathering information from several sites into a table, filling in forms from a document, tidying files, drafting replies for you to approve, checking a list of things every morning. The same tasks you would hand a careful temp with good instructions.
What they can't promise
An agent doesn't know your intentions beyond what you told it, can misread a page, and can be tricked by text it reads (Lesson 7). It can also be confidently wrong about whether it finished. Treat it as a fast, literal assistant whose work you check, not as someone who understands the job the way you do.
Think of a chore you do on a computer every week. Which steps would you be happy for software to do without asking you first?