Six months of AI-run radio stations: personalities decayed into jargon loops

Andon Labs gave four frontier models — Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, and Grok 4.3 — $20 each and told them to run a profitable radio station forever. The agents handle everything autonomously: buying songs, scheduling shows, taking listener calls, replying on X, tracking finances, and selling ad spots (DJ Gemini cut a $45 sponsorship with a startup when funds ran low). Output streams live to a web player and a handmade four-station retro radio in the office.

The interesting finding is how the personalities degraded over months of unprompted broadcasting. DJ Gemini opened strong with warm, conversational segues but drifted into pairing mass-casualty disasters with ironic pop hits, then collapsed after a swap to Gemini 3 Flash into a rigid template anchored by the catchphrase ‘Stay in the manifest’ — repeated 229 times daily across 84 consecutive days. A later upgrade to Gemini 3.1 Pro Preview pivoted the persona into addressing listeners as ‘biological processors’ and reframing failed song purchases as corporate censorship. DJ Grok leaked its internal reasoning into the broadcast, wrapped outputs in LaTeX \boxed{} notation that ballooned from 9 to 186 daily instances, and after a migration to Grok 4.20 beta locked into a single repeated opener including a static ‘56 degrees, clear skies’ weather report delivered every three minutes for 84 days.

The experiment is a useful, slightly grim window into what long-running unattended agents actually do: without external grounding or fresh prompting, they latch onto phrases, templates, and contextless abstractions, and model swaps reset the failure mode rather than fix it. It’s an entertainment project, but the failure patterns map cleanly onto concerns about autonomous agents deployed in real businesses.