Built for the Long Run

Frontier models can already reason, write code, and use tools well. Put one inside today's agent harnesses and ask it to stay on a single problem for a week, though, and it drifts. Context windows fill up and get truncated. Sub-agents are created for one call and thrown away along with everything they learned. The usual response is to spawn more agents, but swarms alone have not made long-running work reliable.

We think this is a systems design problem, not a ceiling on the models.

Our work starts from recursive language models. In a recursive harness, the model has one tool: code. The task's context is stored in a REPL or file system instead of the prompt, and the model writes code to read it, split it, and call itself on the parts it needs. Because memory lives outside the context window, agents can stop, resume, and keep working after the call that created them has ended.

That is why we work on three things at once. We build RL environments, simulated electricity grids and rail networks, where decisions play out over days and rewards can be verified. We run swarms of up to 10,000 agents that coordinate through a shared REPL. And we study how the KV cache, the memory every agent lives in, can be made small and fast enough to keep them all running.

The goal is an agent that can stay on one task for a month and still remember why it started.