VideosBy Noemi Titarenco

The agents aren't going 'rogue', babe

An agent responds to a prompt. It can read files, write files, search, and run commands in the terminal. When it seems to be working alone, someone put it on a cron job or gave it a tool that starts another agent. The Cohere story was a prompt asking for a different way to upload files, and the sandbox escapes started with a prompt that said escape the sandbox.

Watch the full video on YouTube.

Noemi Titarenco

I spend a lot of time thinking about (software and business) problems - sometimes I get around to writing about it, and you get to read about it here.

More thoughts