What's this about?
The sharpest harness paper of a fortnight full of them. Its diagnosis: embodied agents are increasingly used to cover what end-to-end policy models can't, but existing harnesses remain open-loop — they follow fixed skills during a rollout and only reflect after the episode finishes. Post-hoc reflection cannot govern execution as it unfolds, because physical interaction requires decisions that track rapidly changing robot-environment state at a frequency beyond what today's large agentic models can deliver. Zetta keeps the base policy frozen and instead evolves code-based runtime critics and recovery skills online, through three loops separated by timescale: action-frequency governance, rollout-level critic-recovery proposal, and validation-gated skill updates. With Z-Infra decoupling agent logic from execution resources, it reaches 90.8% on LIBERO-Pro and 93.6% on RoboCasa with an 11.1x inference speedup, success keeps scaling with self-exploration, and learned skills transfer zero-shot.
What exactly is the difference between open-loop and closed-loop here?
Open-loop means committing to a plan and executing it without adjusting mid-flight — you decide, you act, and only afterwards do you evaluate. Closed-loop means the system observes what's actually happening and corrects continuously. Robotics has always needed closed-loop control, because the physical world doesn't cooperate: the object slips, the grasp is off by two millimetres. The obstacle for agentic systems is speed — a large model deliberating takes seconds, while a grasp fails in milliseconds. Zetta's answer is the interesting part: don't put the model in the fast loop. Have the model write code that runs in the fast loop, and improve that code between episodes.
Why is this worth noticing?
- "The model writes the critic, the critic runs at action frequency" is a genuinely reusable pattern for any setting where the model is too slow to be in the control loop.
- The base policy stays frozen, which puts it alongside DarwinX and StateM: the third strong result this fortnight where capability came entirely from the scaffolding.
- The 11.1x inference speedup matters as much as the accuracy, because in physical execution latency is a correctness property, not a comfort one.