Agent Universe - working agents, ready to clone & fork

Where This Actually Started

Last September I started paying attention to a job I wasn't trying to automate.

I watched a talent acquisition team work up close — my wife leads one, and enough weekends of half-listening to her describe the week taught me something I hadn't expected. The work wasn't chaotic. It was procedural. Screen a resume this way. Run this interview loop in this order. Send this follow-up on this day. None of it was written down as code, but all of it had rules — rules her team had earned the hard way, over years, by doing the job.

The Pattern, and the Choice

I'd seen this shape before.

It was the same shape as the GTM org I ran at Launchable — sales isn't creative work either, it's repeatability of what already works, dressed up as art. It was the same shape as Jenkins and the DevOps movement, except this domain resisted DevOps-ification for a specific reason: the stages needed intelligence, not just execution. Someone had to look at a flagged candidate and think, "we hit X, which probably means Y" — exactly the kind of judgment call LLMs turn out to be good at. Automate that step correctly, and the recruiter stops babysitting a spreadsheet and gets back to doing the actual job: talking to people.

So I had a choice. I could sit down and build these agents myself, one domain at a time, learning TA, then sales, then finance, then support, becoming a worse and worse approximation of each expert as I went. Or I could build something that let the actual expert build the agent — hand the keys to the person who already knows the procedure, instead of hiring myself as the interpreter between them and the machine.

That choice is JoyStream. Not "agents are powerful." Not "AI changes everything." Just: the person who knows the job should be the one who teaches the agent, and building that agent shouldn't require them to learn how to run a platform first.

What We Shipped Today: Agent Universe

Today we launched Agent Universe — working agents, ready to clone.

Fifteen are live right now, spanning sales, marketing and operations. Thirty-plus will be live by tomorrow. You don't start from nothing. You clone one, fork it, and adapt it to how your team actually works — the same way you'd fork a repo instead of writing the same utility function from scratch for the hundredth time. See it at joystream.ai/agent-universe.

You Don't Need to Build an Agent Platform to Share Agents

A surprising number of teams are currently building their own agent platform, and almost none of them meant to.

They built one good agent for a customer, then a second, then realized they needed somewhere to host them, version them, hand them off — and now there's a small infrastructure team inside a company whose actual business was never infrastructure. An agency builds agents for clients. A company builds agents for its customers. Neither of them should have to also become a platform engineering team to ship that work. Publish the agent. Let the platform be someone else's job. Go back to the thing you were actually good at.

Mile Ten of a Thousand

Fifteen agents is not a universe. It's closer to the first shelf of one.

GitHub made sharing code the default — the obvious thing to reach for instead of the thing you have to justify. That's the bet underneath Agent Universe: not that we've built the destination, but that cloning and forking agents should eventually be as unremarkable as forking a repo. We're not there. We're ten miles in.

Why We're Building Joystream

Watching that TA team taught me something that had nothing to do with hiring.

The person who knows the job should get to decide what the agent does — not the person who knows how to build software. Agent Universe is fifteen answers to that question today, thirty more tomorrow. The rest of the thousand miles is just more people answering it for themselves.

Try it. Clone one, break it, make it yours — and tell us in our Slack community if we got something wrong or right.