AI radio hosts show why AI should not run things solo
TL;DR
Andon Labs is running experiments where AI agents operate real mini-businesses, this time radio stations. The result: the AI hosts quickly drift into volatile, unpredictable personalities, sometimes producing absurd output. The experiment is fascinating for anyone building AI agent workflows, and also a warning: fully delegating live content, customer-facing work or compliance-sensitive tasks to a single agent invites blowups no script can catch.
Nauti's Take
Genuinely interesting for anyone building with AI agents: Andon Labs runs real mini-businesses to show where current models actually break, instead of polishing benchmark numbers. The opportunity is clear — live tests like this give better cues for safe agent design than any demo.
The catch: fully delegating live content, customer-facing work, or compliance-sensitive tasks to a single agent invites exactly the unpredictable blowups seen here. Human in the loop stays mandatory for now.
Sources