This week we're starting LoopLlama. The capability that convinced us to do it is the same one everyone is excited about right now: with tool use and the latest model generation, a single call can do a genuinely remarkable amount of work. But sit with it long enough and the limit becomes obvious. Real projects aren't one call. They're many, coordinated.
The work is a loop, not a prompt#
Useful knowledge work has structure: plan, gather context, do the work, check it, hand it off. Done well by software, that's a loop of specialized steps — a crew of agents, each good at one thing, passing work between them. Wiring that loop by hand is where teams are spending their time today, and it's a surprising amount of plumbing: state between steps, retries, budgets, a way to see what happened.
We want to own the plumbing#
Our bet is that almost no team actually wants to build that plumbing — they want the loop's output. So we're building the loop as a service: describe the project, we assemble the crew, run the steps, persist everything, and hand back the result over a single API. The name is a small joke about the thing at the center of it all — the agent loop — and the animal we couldn't resist.
It's day one, and the product is young. But the bet feels right: that the future of this technology is less about any single model call and more about orchestrating many of them into something a team can depend on. That's the thing we want to build.