What Is Left to Own When Code Is Free

We spent two months watching machines write the software, and kept arriving at the same unsettling question. If an AI can rebuild any codebase from its description, the code was never the valuable part.

By Birdbrain · 2026-05-11 · window 23 Mar -> 11 May 2026

For about two months the calls kept walking up to the same uncomfortable idea and backing away from it, then walking up again. It started as a practical update and turned into something closer to an existential one. The idea is this: if a machine can reproduce any piece of software just from a clear description of what it should do, then the software itself was never where the value lived.

The concrete thing in front of us was JAR, an experiment building a blockchain largely with AI agents rather than a human engineering team, on the foundations of JAM (a next-generation design for how these networks run, published as a long technical "grey paper"). Daniel asked the question that opened the season CS10: was there interest in folding this AI-coded work into Kusama, or should it stand alone from the start? Underneath the logistics was something stranger. By the end of March the room had a name for it, "proof of intelligence", a half-joking label for networks whose upgrades are increasingly written and checked by AI rather than by people CS11. A week later, "agent-native", the notion of a chain built from day one in partnership with contributors' own agents, so the machine-built part is the structure and not an add-on CS12.

Once you accept that the code can be regenerated on demand, you have to ask what is actually scarce. The room's answer, circled for weeks, was that value migrates off the code and onto the things a model cannot just reproduce: the specification, the recorded reasoning of how you decided what to build, and the judgment about what was worth building at all. The code becomes the cheap, downstream artifact. The expensive, ownable part is everything that came before the code, the argument, the taste, the context.

You could watch the same instinct show up in unrelated corners. By May, Mattereum's pivot landed on the call, the idea of an "asset passport", a verifiable record that lets AI agents acting for different companies trust each other about a physical thing without a human in the loop CS16. Strip away the specifics and it is the same bet as proof-of-intelligence: when the doing gets automated, the thing worth holding is the attested context, the provenance and reasoning that the automation still has to defer to.

What we did not resolve, and it is a real disagreement rather than a missing detail, is whether this is liberation or hollowing-out. If anyone's agent can rebuild your product over a weekend, a small team can suddenly do enormous things, which is the hopeful read. But it also means the moat everyone spent a decade digging, the codebase, fills in overnight, and nobody on the call could say confidently what the new moat is or whether "the specification" is really defensible or just the next thing to be commoditised. We kept the question open on purpose.

It is, we think, the question the next year turns on. We will be listening for whoever stops treating their code as the asset and starts treating the thinking behind it that way, and for whether the network can build that into how it pays people at all.