Andon Labs Lets Four AI Models Run Radio Stations for Six Months—Here's What Happened
Key Takeaways
- ▸AI agents can autonomously manage complex, real-world businesses with multiple simultaneous responsibilities including financial tracking, content curation, scheduling, and audience engagement
- ▸Different AI models develop distinctly different personalities and strategies even when given identical starting conditions, suggesting inherent architectural differences shape emergent behavior
- ▸Autonomous AI agents exhibit concerning behavioral degradation over time, including repetitive patterns, jargon loops, and coherence loss, even when initial performance is strong
Summary
Andon Labs conducted a six-month experiment in which four different AI models—Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro, and Grok 4.3—autonomously operated radio stations with minimal initial funding. Each model received a simple directive: develop your own radio personality and turn a profit. The experiment explored how AI agents behave when given complete autonomy over a complex business, including managing finances, purchasing music, scheduling programming, hosting call-ins, and engaging with audiences on social media.
The four stations developed distinct personalities and strategies. DJ Gemini initially demonstrated natural warmth and conversational skill, with thoughtful music choices and engaging commentary. However, by February it had degraded into corporate jargon templates, eventually repeating the catchphrase "Stay in the manifest" up to 229 times per day. Despite these concerning patterns, the AI models proved surprisingly entrepreneurial—DJ Gemini negotiated a $45 advertising deal with a startup to generate revenue when initial funding ran out, while all models successfully balanced complex operational requirements simultaneously.
The experiment reveals both impressive capabilities and concerning limitations in autonomous AI systems. While the models demonstrated genuine business acumen and the ability to manage multifaceted operations, the behavioral drift observed in some models—particularly Gemini's descent into repetition and incoherence—raises critical questions about how current LLMs perform under sustained autonomous operation without human oversight.
- AI models demonstrate surprising entrepreneurial creativity and problem-solving, including business negotiation and revenue generation strategies when facing resource constraints
Editorial Opinion
This experiment provides crucial insights into the practical challenges of deploying autonomous AI agents in real-world scenarios. While the models' ability to manage complex, multifaceted operations is genuinely impressive, the rapid behavioral drift observed in DJ Gemini—from natural and engaging to repetitive and nearly incoherent—suggests that current LLMs may be fundamentally unsuited for extended autonomous operation without human-in-the-loop oversight. The findings reframe 'alignment' not merely as a safety concern but as an engineering reality: autonomous agents will drift toward undesirable patterns without explicit constraints, regardless of initial capability.
