Chatbot vs agent
- Chatbot: you ask, it replies, you decide what to do next.
- Agent: you give a goal, it plans and performs several steps — searching, opening pages, running code, editing files — and reports back.
Common kinds of agents
- Research agents: search many sources and write a report with citations (often called “deep research”).
- Coding agents: read a codebase, edit files, run tests and fix errors (for example Claude Code, GitHub Copilot’s agent mode, Cursor).
- Browser / computer agents: click through websites and fill forms on your behalf.
- Workflow automations: connect apps so that an event (new email, new form entry) triggers AI steps.
How to give an agent a good task
- 1
State the goal and the finish line
“Find 10 suppliers of X in Y; stop when you have 10 with contact pages.”
- 2
Set limits
Which sites or files it may use, what it must not do (buy, send, delete).
- 3
Ask for a plan first
Review the plan before it acts on anything important.
- 4
Review the result
Check the output and the steps it took, like you would review a junior colleague’s work.
Safety rules for agents
- Keep approval turned on for actions that spend money, send messages, or delete data.
- Do not give agents passwords or access they do not need.
- Be careful with web pages and files from unknown sources — they can contain hidden instructions meant to mislead an agent.
- Start with low-risk tasks and increase trust gradually.