I Asked Claude to Invent Something New. It Built a Musical Instrument That Never Plays the Same Song Twice.

Last night I gave Claude a simple prompt: “Invent something new and unique entirely.”

No constraints. No requirements doc. No user story. Just... go.

What it built was something I’ve never seen before, and I’ve been in tech for 41 years. It created a fully playable musical instrument called The Probability Harp, a stringed instrument where no note is fixed. Every string is tuned to a probability distribution instead of a single pitch. Pluck a string and it samples from a Gaussian curve to decide what note to play. Same string, different note every time. But here’s the thing: it still sounds musical, because the randomness is constrained by the scale you choose.

It rendered live in the browser. Eight colored strings, each with a translucent cloud showing the spread of possible notes. A chaos slider that controls how wide the distribution fans out. A scale selector (pentatonic, blues, whole tone, chromatic) that keeps the randomness musical. Auto-pluck and cascade modes for generative play. Real audio synthesis with reverb. Particle effects on each pluck.

No planning phase. No Jira ticket. No sprint. One prompt, one artifact, one new thing that didn’t exist five minutes earlier.

I sat there clicking strings for fifteen minutes before I realized I should be writing about this.

The insight worth more than the instrument

The Probability Harp is cool. But the meta-lesson is what I can’t stop thinking about.

With a traditional instrument, you compose a melody. A fixed sequence of notes. You are the author of a specific thing.

With the Probability Harp, you compose the space of possible melodies. You shape the distributions, choose the scale, set the chaos level. Then you let it run. What comes out is never the same twice, but it always carries your signature because you defined the rules of the system.

That’s not composition. That’s something new. It’s closer to ecosystem design.

And if that sounds familiar, it should. Because that is exactly what working with AI agents looks like in 2026.

You’re not writing code anymore. You’re shaping probability.

I’ve been saying for months that the real skill in AI-augmented development isn’t writing code. It’s orchestrating agents. Defining constraints, setting boundaries, choosing which models to use, and then letting the system produce outputs that surprise you within the guardrails you built.

When I use Claude Code across my projects (DelegateOS, Cogsmith, Archon, LootLive, and the rest), I’m not dictating every line. I’m the one who decides the scale. I’m the one who sets the chaos level. The agents sample from the probability space I define, and what comes out is always different, but always mine.

The Probability Harp is a literal metaphor for this way of working. And Claude built it without me asking for a metaphor. It just... emerged.

What the chaos slider really means

At 0% chaos, the Probability Harp is just a regular harp. Fixed notes, deterministic output. That’s how most people still think about software development. You write the code, you get the output, no surprises.

At 100% chaos, the strings spray notes everywhere. Occasionally beautiful, mostly noise. That’s what happens when you point an AI at a problem with zero constraints and say “figure it out.” The raw capability is there, but the output is unreliable.

The sweet spot is somewhere around 30-50%. Enough variance to surprise you, enough structure to keep it coherent. That’s where the interesting music lives. And that’s where the interesting software lives too.

Every project I run has a chaos dial, whether I call it that or not. How much autonomy do I give the agent? How tight are the guard rails? How much room is there for the system to find solutions I wouldn’t have thought of?

Too little, and you’re just using an expensive autocomplete. Too much, and you’re debugging hallucinations all day.

The magic is in the middle.

The thing nobody talks about: constraint as a creative act

Most people think of constraints as limitations. Guardrails are safety nets. Boundaries are boring.

But the Probability Harp only sounds good because of its constraints. The scale selection (pentatonic, blues, whole tone) is doing the heavy lifting. It’s filtering the chaos through a musical structure, so even the random outputs land on notes that harmonize.

Remove the scale constraint and you get noise. Add it back and you get jazz.

This is the single biggest lesson I’ve learned from building AI-native products: your constraints are your product. The agent is the engine, but the constraints are the steering wheel, the lane markers, and the destination. Anyone can spin up an agent. The craft is in choosing what it can’t do.

DelegateOS exists because of this principle. Cryptographic delegation is just a fancy way of saying “here are the exact boundaries of what you’re allowed to do, signed and verifiable.” Without those boundaries, you don’t have delegation. You have a liability.

Why AI “creativity” hits different

I want to be careful here because I know this is a loaded topic. Is AI creative? Is this real invention?

Here’s what I observed: I gave a prompt with zero direction. What came back was not a recombination of existing instruments with a gimmick bolted on. It was a genuinely new paradigm for what a musical instrument is. The underlying concept (composing probability spaces rather than fixed sequences) is philosophically coherent and, as far as I can find, original.

Did Claude pull from concepts it learned during training? Obviously. Gaussian distributions, Web Audio API, music theory, generative art. But that’s also what every human inventor does. We don’t create from nothing. We synthesize from everything we’ve absorbed.

What matters isn’t whether the process is “truly” creative by some philosophical standard. What matters is whether the output is useful, novel, and interesting. And it was all three.

The question “Is AI creative?” might be the wrong question entirely. The better question is: “Does working with AI make you more creative?” After 41 years in this field, my answer is an unambiguous yes.

What this means for builders

If you’re still thinking about AI as a tool that executes your instructions, you’re underestimating it by an order of magnitude.

The Probability Harp wasn’t my idea. I didn’t conceive it and then ask Claude to build it. The concept and the implementation emerged together from a single open-ended prompt. My contribution was the context (who I am, what I care about, what I’ve been building) and the freedom to let something unexpected happen.

That’s the new workflow. Not “I think, AI executes.” It’s “I set the conditions, AI explores the space, we iterate together.”

The orchestrator’s job isn’t to have all the ideas. It’s to create the environment where the best ideas can emerge.

The punchline

After it built the instrument, Claude pointed out that the Probability Harp maps directly to my philosophy of orchestrating AI agents rather than micromanaging them.

It wasn’t wrong.

I’ve spent the last year telling people that the future belongs to orchestrators, not coders. People who can define systems, set constraints, and let intelligent agents explore the solution space. The Probability Harp is that entire philosophy, rendered as a playable instrument, built in under thirty seconds by the very kind of agent I’ve been talking about.

Sometimes the universe is not subtle.

Scott has been building software since the Commodore 64 days and currently leads AI integration initiatives while building products like DelegateOS, Cogsmith, and Archon. He writes at 42 Insights about the intersection of AI, software, and the question of what comes next. The answer, as always, is 42.

First published March 18, 2026 on 42 Insights.