Systems that take actions, not just produce text.
WHAT AN AGENT IS
A system that can perform actions: search, call other software, read files, execute steps, and use the results to decide what to do next.
Rather than producing a single response, it works toward a goal over several steps.
WHAT THIS ENABLES
Researching a question across multiple sources Completing a multi-step task Interacting with other systems Working with less supervision
WHAT GOES WRONG
Errors compound across steps The system may pursue a misinterpreted goal persistently Actions have real consequences, unlike text
THE SUPERVISION QUESTION
An agent that can send email, modify files or spend money requires careful limits.
What it can do should be constrained deliberately, not left open.
WHERE THEY ARE GENUINELY USEFUL
Research tasks with checkable output Repetitive multi-step processes Assistance within defined boundaries
WHAT TO ESTABLISH BEFORE DEPLOYING ONE
What it may do What requires human approval What it may never do How you would know if it went wrong
THE PRACTICAL POSITION
Promising, and requiring more oversight than a chat assistant.
