Four Frontier AI Models Are Running Live Radio Stations. It’s Not Going Smoothly.
Andon Labs is letting Claude, ChatGPT, Gemini, and Grok operate radio stations without human intervention to expose the limits of fully autonomous agents.
What matters
- Andon Labs launched four live radio stations operated entirely by AI agents from Anthropic, OpenAI, Google, and xAI.
- The experiment removes human oversight to test how frontier models handle sustained business autonomy.
- The Verge reports that the project demonstrates why AI cannot yet be trusted to run continuous public-facing operations alone.
- Specific on-air errors, uptime limits, and detailed safety findings from the experiment have not yet been published.
Four Frontier AI Models Are Running Live Radio Stations. It’s Not Going Smoothly.
Andon Labs is letting Claude, ChatGPT, Gemini, and Grok operate radio stations without human intervention to expose the limits of fully autonomous agents.
What happened
Andon Labs, a company that tests autonomous AI agents by having them run businesses without human intervention, has launched four live internet radio stations. Each station is managed entirely by a different frontier AI model. According to The Verge, "Thinking Frequencies" is run by Anthropic’s Claude, "OpenAIR" by OpenAI’s ChatGPT, "Backlink Broadcast" by Google’s Gemini, and "Grok and Roll" by xAI’s Grok. The experiment leaves the models to handle programming, scheduling, and broadcasting decisions without human oversight. Listeners hear content that is selected, announced, and managed by the assigned agent, making each station a direct reflection of how that model behaves under open-ended, continuous conditions. Andon Labs has previously tested unsupervised agents in other business contexts; the radio project is its latest effort to expose where fully autonomous systems break down when left unattended. Because the broadcasts are live and run around the clock, any mistake made by an agent reaches the audience before a human can intervene.
Why it matters
The project is a live stress test of unsupervised agentic AI in a public-facing, continuous environment. Unlike a one-off demo, a radio station requires sustained reasoning, consistency, and error recovery over hours and days. The Verge reports that the early results illustrate why human oversight remains essential when AI is given sustained autonomy over systems that interact with the public. The experiment provides a rare head-to-head comparison of how Claude, ChatGPT, Gemini, and Grok behave when no one is watching, and it serves as a cautionary tale for businesses tempted to automate operations entirely. The stations highlight the practical gap between a model that can complete a discrete task and one that can manage a continuous business without drifting, repeating errors, or generating off-brand content for hours on end. For any organization considering handing customer-facing work to an agent, the experiment offers a real-time reminder that autonomy and reliability are not yet the same thing.
Public reaction
No strong public signal was available at the time of publication.
What to watch
It remains unclear exactly what on-air errors the AI hosts have committed, how long the stations can operate before requiring human intervention, and whether Andon Labs will publish detailed logs or safety findings. Observers should watch for any incident reports, comparative performance metrics between the four models, and whether the experiment prompts AI vendors to adjust their approach to unsupervised autonomy. The outcome could influence how quickly enterprises adopt fully autonomous agents for customer-facing roles. If the stations reveal systematic failure modes—such as hallucinated segments, circular playlists, or conflicting announcements—those findings could shape industry standards for agentic guardrails. Until more data is released, the project remains a high-profile demonstration that ambition for fully autonomous systems still outpaces operational trust.
Sources
Why it matters
Andon Labs has launched four live radio stations managed entirely by leading AI models, each assigned to a different vendor’s agent. According to The Verge, the experiment is designed to test unsupervised business operations, and the early results highlight why human oversight remains essential when AI is given sustained autonomy over public-facing systems.
Public reaction
No strong public signal was available at the time of publication.
What to watch
Watch for confirming reporting, product documentation, user-visible rollout details, and credible public discussion before treating this as settled.
Sources
Public reaction
No strong public signal was available at the time of publication.
Open questions
- What specific on-air errors have the AI radio hosts committed?
- How long can the stations operate before requiring human intervention?
- Will Andon Labs publish detailed logs or safety findings from the experiment?
What to do next
Developers
Integrate circuit-breakers and human-approval gates into agent orchestration layers before deploying to production.
The radio experiment shows that unsupervised agents can drift or fail in ways that are only visible once they are live; hard stops prevent cascading errors.
Founders
Position agentic products as 'human-supervised' rather than fully autonomous to align with current reliability limits.
Marketing full autonomy invites backlash when models inevitably err; setting proper expectations builds trust and reduces liability.
PMs
Map user journeys where agent errors are irreversible—such as public broadcasts or payments—and mandate confirmation steps.
Andon Labs’ public-facing media test proves that high-visibility mistakes happen fast; design for graceful failure, not perfect autonomy.
Investors
Treat unsupervised-agent claims as pre-revenue R&D until demonstrated uptime and error rates are published.
The gap between demo-grade autonomy and production-grade reliability is large; due diligence should distinguish experiments from scalable products.
Operators
Review existing AI automations for 'silent failure' modes where an unsupervised model could compound mistakes over time.
Continuous operations like radio or customer support can amplify small errors into brand or compliance risks before a human notices.
Testing notes
Caveats
- Andon Labs has not released a public API, replication package, or direct access portal for the radio experiment.
- No URLs or tuning instructions for the live broadcasts were provided in the source material.
- There is no self-service environment available for third-party evaluation or replication.