My long-term question is whether models can develop capabilities beyond human performance when feedback comes directly from the environments in which they act.
I work with LLMs to study how agents act, receive feedback, and adapt in real textual environments. The harness is the system that connects a model to its environment, shaping what it can observe, do, and learn from.
hxlog turns classic books into structured maps of concepts, relationships, and learning paths. Rather than serving as linear reading notes, it is designed as a knowledge structure that both people and agents can use to navigate and teach the material.
My personal research workspace, where I experiment with using agents for literature discovery, domain learning, idea review, experimental design, and execution. The goal is to understand where they can meaningfully support scientific work.
An early-stage experiment in self-improving agent harnesses. I will share more once the design is stable enough to evaluate.
-
Backend Engineer · NeuroXess
A life-science company focused on the research and clinical application of implantable, flexible brain-computer interfaces.
I also enjoy contributing to open-source projects around AI agents—submitting PRs, fixing issues, and extending useful tools. If you are working on agent infrastructure, evaluation, or related tools, I would be glad to exchange ideas or collaborate.
Computer Science student at Shenyang Agricultural University, 2023 cohort.