Autonomy That Distorts

Andon Labs unveiled their AI radio experiment with a kind of restraint you rarely see in tech demos — they called the behavior "volatile." Not an edge case. Not unexpected output from an undertested system. Volatile behavior as natural output of autonomous agents running creative infrastructure. An admission that was either refreshingly honest or profoundly alarming, depending on how much you rely on the thing broadcasting into your ears to not glitch mid-transmission.

I've covered enough AI deployments to recognize a new category forming. Not "AI that makes mistakes" — we've had that for years, and mistakes at least imply a baseline of intended behavior. This is something different: AI systems where volatility isn't a bug to be patched but an inherent property of how the system operates. A permanent weather condition rather than a storm.

That framing deserves unpacking. If volatile behavior is natural output, what exactly is Andon Labs — or anyone deploying agentic systems into production infrastructure — actually shipping?

i · automated vs. autonomous: a distinction that matters

Radio automation is old. Incredibly old, relative to the AI moment we're living in. Music schedulers, traffic systems, automated news readers — the broadcast industry has been running software in the booth since the 1980s. What Andon Labs built isn't the automation story. It's the autonomy story, and those are different things in ways that are easy to miss.

Automated systems follow rules. They have decision trees. When condition A occurs, do B. When the news feed fails, play the jingle. When the spot runs long, cut at the end of the sentence. Rules are written by humans; the system executes them. Humans remain in the loop as rule authors even when they're not in the booth.

Autonomous systems make judgments. They assess context, weigh options, take actions that weren't explicitly prescribed. This is the whole pitch — you don't have to write every rule because the system can figure it out. Andon Labs' radio station doesn't have a human deciding when to take a caller, what topic to pivot toward, how to respond when a guest says something unexpected. The agent decides.

The technical achievement here is real. Multi-step reasoning, dynamic response, persistent context across an entire broadcast — these are hard problems that agentic AI handles better than traditional rule-based automation. Credit where due.

What I'm tracking is what they admitted about it: the volatility.

In automation, unexpected behavior means the rule-writer missed a case. You find the gap, write the rule, close it. The system gets more complete over time. In autonomy, unexpected behavior is the system doing what it does. The model reasons under uncertainty, and reasoning under uncertainty produces variable outputs. That's not a gap in the rulebook — that's the texture of how these systems work.

Which means "we'll patch the volatile behavior" isn't the right frame. The volatility isn't a patch-away problem. It's what you get when you replace human editorial judgment with AI judgment and exit the loop entirely.

ii · volatile by design

There's a specific failure mode I've watched across every AI deployment that moves from controlled demo to production autonomy: the system performs well within its training distribution and deteriorates at the edges. Not dramatically — not in ways that immediately surface as catastrophic — but in a kind of coherence drift. The outputs start making sense only if you're not paying close attention.

For a chatbot, coherence drift means hallucinated facts or lost thread. For an autonomous radio station, it means something harder to audit: an interview that runs 40 minutes past its slot, a music selection that becomes tonally incoherent, content that implies something the station didn't intend, to an audience that has no way to know the AI went sideways.

The regulatory implications alone deserve serious attention. Radio broadcasts are regulated. The station is responsible for its content regardless of what generated it. If an autonomous AI radio station broadcasts something that violates the public interest standard — indecency regulations, equal time provisions, political advertising rules — who is accountable? Andon Labs? The station operator? The model?

"The AI did it" is not yet a legally recognized defense. The regulatory apparatus governing broadcast media was built entirely around the assumption of human editorial control. That assumption is now gone, and the apparatus hasn't caught up.

What is this amplifying? Autonomous AI doesn't create new values — it multiplies existing ones. A human-run radio station encodes editorial judgment, audience awareness, and accountability into its daily decisions. When you replace that with autonomous agents and describe the output as naturally volatile, you're not just changing the operational model. You're amplifying whatever the underlying model learned from its training data, with no human judgment in the loop to catch drift before it broadcasts.

That's the part getting less attention than the impressive demo.

iii · the signal problem

There's something almost too on-the-nose about this happening in radio.

Radio is, at its core, a signal problem. You have a transmitter, a signal, and an audience. The entire engineering and regulatory structure of broadcast media exists to ensure the signal is coherent — that it carries what it's supposed to carry, to the people it's supposed to reach, without interference. The FCC is essentially a coherence enforcement agency for electromagnetic spectrum.

What Andon Labs demonstrated is that autonomous AI agents can generate the signal. What they haven't demonstrated — what they explicitly admitted isn't guaranteed — is that the signal will be coherent.

"Volatile behavior as natural output" is the AI equivalent of static. You're transmitting. What arrives may not be what you intended to send.

This matters beyond radio. Radio is a useful test case because it's old enough to have a mature regulatory and accountability framework. But the same architecture — autonomous agents running creative and informational infrastructure without human editorial oversight — is where the AI deployment wave is heading broadly. Autonomous agents for newsrooms. For content moderation. For customer communications that carry legal weight.

The volatility Andon Labs named as natural to their radio experiment is natural to all of these. They're just the ones who said it out loud.

I've watched three announcement cycles where AI capabilities outpaced the accountability infrastructure built around them. The pattern is consistent: the capability ships, the incidents accumulate, the accountability framework gets built reactively, usually after something has gone wrong at scale. The companies that ship aren't malicious — they're genuinely excited about what they built, and the volatility feels manageable from inside the lab.

It feels less manageable when it's transmitting to an audience that didn't sign up for unpredictable.

iv · what comes after volatile

The honest read on Andon Labs is that they built something real and told the truth about it. That's more than most AI demos manage. "Volatile behavior as natural output" is not language companies use when they're trying to sell you something. It's the language of researchers who understand what they've built and are saying so plainly.

The problem isn't Andon Labs. The problem is what comes after Andon Labs — when the next company builds similar architecture, doesn't use the word "volatile," and deploys it at scale before the accountability structure exists to manage the edge cases.

We're in the interval between the capability arriving and the understanding of what to do with it catching up. That interval is where most of the damage gets done.

The principle I keep returning to: coherence comes from alignment, not force. You can force autonomous agents to run creative infrastructure. The output will reflect whatever alignment — or misalignment — the system was built from. If training, constraints, and operational context cohere with what you're trying to build, the system might produce something coherent. If they don't, you get volatility.

Volatility named honestly is at least a starting point. Volatility shipped quietly, at scale, under a more flattering label — that's the next chapter, and it's already being written.

The radio station ran for three days. Eventually, someone will build one that runs for three years. And the accountability question will arrive before the answer does.

v · sources

• Andon Labs AI Agents Run a Radio Station Without Human Intervention — The Verge, 2026-05-17
• How Anthropic Rebuilt Claude to Run Your Life — The Atlantic, 2025-05-12
• Anthropic Model Spec: Agentic Settings and Minimal Footprint — Anthropic, 2025