Let the Agent Wander the App First, and Kimi K3 Beats GPT-6
RSIAgent, submitted September 14 at https://arxiv.org/abs/2609.15364 by Sibo Zhu, Shicheng Fan, Xinyue Wang, Wenyi Wu, Kun Zhou and Biwei Huang, is training-free, which is the first thing worth saying about it. No fine-tuning, no gradient updates, no weights touched. It just lets the agent explore an unfamiliar environment before you ask it to do anything, and keeps what it learned.
The strategy is broad-then-deep. Parallel broad self-exploration maps out the structure of the environment — what screens exist, what a button does, where things live. Then focused deep exploration goes after the hard cases the broad pass flagged. Three roles coordinate it: a curriculum agent deciding what to explore, an actor doing it, and a verifier checking whether what came back is true. The output is an environment-specific memory, and then the memory is frozen. It gets reused on downstream tasks with no parameter updates at all.
The headline result: this lets Kimi-K3 and GLM-5.3 outperform frontier closed models including GPT-6 on OSWorld-v2 and Agent's Last Exam. Take that with the usual salt for a paper making its own comparison, but take the structure seriously. An open model plus a good map of the environment beat a better model with no map. That is a statement about where the missing capability actually lives, and it is not in the weights.
This is the same lesson arriving from a third direction. [Show-Harness argued robots don't need a new model, they need a better interface](https://clauday.com/article/66aa6315-7239-4f06-a8aa-503c9c8e0ef3). [Thirteen clever RL data recipes all failed to beat random selection](https://clauday.com/article/3d4aead7-c918-405b-91a6-55524e112d57). Now a training-free exploration pass closes a frontier-model-sized gap. The pattern across all three is that we have been attributing to model quality what actually belongs to environment knowledge, and the cheapest wins available right now are in the harness.
Practically, if you run agents against an internal tool — your own admin console, your own CRM, some ancient enterprise UI — this is directly usable today. Let it wander for an hour, freeze what it found, ship the memory as an artifact next to the prompt. The interesting unanswered question is staleness: a frozen map of a UI that ships weekly is a map that is wrong by Thursday, and the paper does not tell you how to notice.
← Back to all articles
The strategy is broad-then-deep. Parallel broad self-exploration maps out the structure of the environment — what screens exist, what a button does, where things live. Then focused deep exploration goes after the hard cases the broad pass flagged. Three roles coordinate it: a curriculum agent deciding what to explore, an actor doing it, and a verifier checking whether what came back is true. The output is an environment-specific memory, and then the memory is frozen. It gets reused on downstream tasks with no parameter updates at all.
The headline result: this lets Kimi-K3 and GLM-5.3 outperform frontier closed models including GPT-6 on OSWorld-v2 and Agent's Last Exam. Take that with the usual salt for a paper making its own comparison, but take the structure seriously. An open model plus a good map of the environment beat a better model with no map. That is a statement about where the missing capability actually lives, and it is not in the weights.
This is the same lesson arriving from a third direction. [Show-Harness argued robots don't need a new model, they need a better interface](https://clauday.com/article/66aa6315-7239-4f06-a8aa-503c9c8e0ef3). [Thirteen clever RL data recipes all failed to beat random selection](https://clauday.com/article/3d4aead7-c918-405b-91a6-55524e112d57). Now a training-free exploration pass closes a frontier-model-sized gap. The pattern across all three is that we have been attributing to model quality what actually belongs to environment knowledge, and the cheapest wins available right now are in the harness.
Practically, if you run agents against an internal tool — your own admin console, your own CRM, some ancient enterprise UI — this is directly usable today. Let it wander for an hour, freeze what it found, ship the memory as an artifact next to the prompt. The interesting unanswered question is staleness: a frozen map of a UI that ships weekly is a map that is wrong by Thursday, and the paper does not tell you how to notice.
Comments