Nexus
Nexus finds everyone who owns a piece of a song and asks each of them for a yes, so a label can offer that song for remixes.
One percent is enough to say no.
That's the problem I designed Nexus around.
Before a label can offer a song for remixes, everyone who owns a piece of it has to say yes: every songwriter, each writer's publisher and the companies that administer those publishers. It's a permission slip with a line for every owner.
For a hundred songs, that's an operations job most catalog teams aren't staffed to run. (My founding document for Nexus starts with the question labels ask once they like the pitch: do the publishers have to approve too? They do.)
Who wrote the song? Who publishes each writer? And who gets to say yes?
The first two are in public registries, kept by people and often behind the latest deals. The third is written into each writer's deal, and those deals are private.
AI agents go after the first two. They take a recording's code and search the registries and streaming catalogs for the song behind it, its writers, and the publishers and administrators who control each part.
One registry listed a colleague of mine under three spellings. "I don't get it," I said in May. "No wonder there's no system that does this."
The agents merge those variants into one record, note which source said what and give each record a confidence score, so a person can check the work.
Then Nexus asks. It writes to publishers and administrators first, because they know who else has to sign. The first email is a fixed template, sent from the label's own mailbox. A model reads each reply, quotes the sentence that matters and suggests one of five next moves.
A person presses the button.
Once every owner has signed, the song is ready to list on Vers1ons. When an owner can't be found or can't approve, Nexus says so.
Where the AI stops
Replies come from outside the company, so Nexus treats them as untrusted. A model can suggest adding a new owner, and only a person's click makes it real. (The reply I worry about most is a stranger claiming a share.)
To test the part that reads replies, an engineer wrote 17 made-up email threads. An autoreply quotes a typical wait of 10 business days, and Nexus passes only by doing nothing. One publisher sends a cheerful yes, for a different song. Nexus has to leave that yes alone and ask about the right one.
Plain code checks each signed document's text against what Nexus sent, and reads the date.
A permission slip signed before it was drawn up, or dated after it came back, doesn't count.
A music-publishing expert set my expectations in June: Nexus can be right 75 to 80% of the time. When a writer's deal ends, the old publisher can keep collecting for 12 months with no right to approve anything, and the registry still lists it. One more reason Nexus asks.
FINAL, then TRULY FINAL
Nexus is one part of the system I designed for Vers1ons. I built its first working version over one weekend in May 2026, from the database to the agents to the PDFs.
At 9:44 Sunday morning, the commit for the last of nine planned agent skills said FINAL. At 10:11 I reverted part of it, because it broke the deploy. At 12:16 the real last one landed: TRULY FINAL.
Nine days later I told the company, "It is alive now. It's not necessarily working well, but it's alive."
In June, after a session with that expert, I built the first outreach flow and had it in production three days later. Opening its approval link changed nothing. Only pressing the button counted, because corporate mail scanners open every link, and a scanner shouldn't get to sign a permission slip.
In July another engineer wired in the first reply-reading loop. In August the engineer who wrote those tests rebuilt the flow around email threads and rounds of outreach, and he has run it day to day since. He taught Nexus to recognize a signed document.
His first Nexus commit, 12 days after my TRULY FINAL, was titled "fix local dev."
Fair.
Stack
- Languages
- TypeScript
- Models and agents
- Claude (Sonnet and Haiku)LiteLLMPi agent SDKBAML for reading replies