Beyond chat.
THE DISTINCTION
A chat assistant produces a response.
An agent takes actions: searching, reading files, calling other software, and using results to decide what to do next.
WHAT THIS ENABLES
Multi-step research Completing a task rather than describing it Interacting with other systems
WHAT MAKES THEM DIFFICULT
Errors compound across steps A misinterpreted goal is pursued persistently Actions have consequences that text does not
THE SUPERVISION QUESTION
An agent that can send email, modify files or spend money needs explicit limits.
WHAT TO ESTABLISH BEFORE USING ONE
What it may do What requires approval What it may never do How you would know if it went wrong How to stop it
WHERE THEY WORK WELL
Research tasks with checkable output Repetitive multi-step processes within clear boundaries
WHERE THEY DO NOT
Anything consequential without approval Open-ended goals
THE HONEST ASSESSMENT
Promising, improving, and requiring more oversight than people give them.
