Most agent systems throw away their most suitable training data after every run.

A task finishes, and the useful part disappears with it: which model got routed where, which tool was called, what came back, where the agent failed, and how it recovered.

And now, TokenRhythm released NeoHorse-1, a 4B/9B model family post-trained on agent execution traces and outcomes.

So it is trained on what an agent actually did, including its tool calls, mistakes, and recoveries, then that experience is fed back into training so the next model can perform better.

NeoHorse-1 is an early engineering validation of RSI through two connected loops: Data-RSI + Model-RSI. Both the NeoHorse-1 checkpoints start from Qwen3.5. The training loop around them uses structured execution trajectories generated inside

TokenRhythm's OpenSquilla Harness, capturing signals such as routing decisions, tool calls, failures, recovery steps, and outcomes. The updated model goes back into OpenSquilla to generate the next round of trajectories.

This is a really solid direction because, AI doesn’t automatically improve just because it uses tools. The key is whether execution experience can be captured, evaluated, and turned into useful training data.