Models are ready.
Harnesses aren't.
Today's agent systems are ephemeral. Context windows fill up, sub-agents vanish when their call returns, and spawning thousands more still doesn't hold a task together for a month. The gap between what models can do and what they actually do is a systems problem.
We're closing it with a recursive harness, inspired by recursive language models, and the RL environments to train models inside it. The goal: agents that work on long tasks without losing memory.
Recursion over context, not a bigger window
Inspired by recursive language models, our harness has one tool: code. Task context lives in a persistent REPL and file system, never crammed into the prompt. The root agent writes code to inspect it, calls itself on the pieces it needs, and gets results back as variables instead of tokens.
Sub-agents persist. They can pause, resume, and outlive the root call. We train models inside this harness with verifiable RL environments, so the strategy that solves a task at the top is the same one each sub-agent uses on its smaller piece.
