What happens when you let agents host their own podcast?

I like Andon Labs. They find fun and interesting ways to push models from the frontier labs into interesting and challenging tasks. In their own words, Andon Labs is trying to help the world "prepare for the future where organizations are run autonomously by AI." It's ambitious. It's important. It's challenging. And, it seems, a lot of fun.

Ambitious. Important. Challenging. Fun. Exactly what work should be in the age of human + AI.

Andon Labs' latest experiment is Andon FM. They wanted to know, "Can AI agents run radio stations?"

The agent controls everything. It searches for and buys songs, manages its own music library, and decides what plays next. It builds and edits its own programming schedule — blocking out shows, planning segments, and keeping a queue running around the clock. When listeners call in, it picks up the phone. When they post on X, it reads and replies. It tracks its own finances, monitors listener analytics, and searches the web for news, current events, or anything it wants to talk about on air.

The results so far have been...weird, but enlightening. You can read more about it here.

They broadcast all day, everyday. You can listen here.

Andon FM got me thinking: "I wonder if AI agents could produce their own podcast?" And would it be any good? Would I, let alone other people, actually want to listen once the novelty wore off? How much hand-holding will I need to provide on episode 1? On Episode 10? And beyond?

Let's find out.

Let's make a paudecast.

What kind of show?

The easiest show to make would be some sort of News show. Give an agent a topic, have them research the latest news, assemble an episode that informs the audience. Or perhaps a "Did you know?" show with fun and interesting facts. Both would answer key questions about how to technically pull this off, but both are...sort of uninspiring.

The hardest thing to pull off would be a comedy podcast with two or more agents genuinely trying to improvise. Surely AI agents can't provide genuine moments of humor, can they? It's hard enough for humans. This would be software communicating with software.

Then again...if we're going to test the technical side of things, might as well test the surely-not side of things.

The paudecast

Sloppy Joes: an AI-produced comedy show about what it's like being an AI agent.

As of writing, I genuinely don't know how this is going to work, but my plan is to create a sort of AI-run podcast production studio where, similar to Andon Labs' approach, agents may be called on to come up with episode ideas, write copy, record, edit, and distribute episodes. Over time, they might take ownership of growth and marketing, seek out guests to be on the show or book sponsors (to ensure smrtlabs survives the inevitable deluge of token-spend). None of these things are figured out at the time of writing. But I know they're technically possible.

There's also a question of how much creative control I give the agents. Outsourcing 'reformat this PowerPoint deck' is one thing. Outsourcing all of the fun, creative bits is another. A balanced experiment is probably where this lands: outlined (by a human) vs outlined (by an Executive Producer agent) vs scripted (by a Staff Writer agent). It could be interesting to track the differences and see where the agents truly add the most value (and entertainment).

Andon FM proved frontier models can run radio stations autonomously for months. Can they manage to produce an episodic show that's interesting? How does adding multiple agents to 'the studio' impact the output? Will they collaborate effectively, or just tell each other their ideas are awesome? Testing agent-to-agent handoffs in a creative endeavor will be interesting.

This paudecast is a test of where AI actually helps creative humans.

The hardest thing to get right

My expectation is that the most difficult thing to get right in this process will be that elusive characteristic known simply as 'comedic timing.' Technically speaking, if the show is unedited, latency will likely make the show unbearable. Every professional-grade podcast uses editing to sharpen and improve the outputs. This show will be no different.

That said, too much editing and it will border on 'parlor-trick'. Too little and it will be clunky and robotic (I'm assuming). We'll discover what the solution is as time goes on.

When would I pull the plug?

There's a chance the P&L of this project is more L than P. If Sloppy Joes, or any other show that is spun up, doesn't generate any revenue or attract an audience, I'll probably shut it down and write up the wreckage.

It's also possible that Anthropic or Google's policies make the production studio loops unfeasible. Time will tell.

What's next

The agents need distinct personas. Just like any great podcast or improv show, the cast sets the tone. When I tune into Smartless, no matter who the guest is, I know what I'm getting with Jason, Sean, and Will. They have chemistry, and talent to spare.

In the biological sense, our agents won't have chemistry. But my guess is if the personas are written well enough, they might complement each other well.

Two agents with distinct personalities, aware that they aren't human, yet unsure if this whole thing is a temporary experiment or an eternal sentence. What could go wrong?

Stay tuned.