Hexo Labs Published SIA — an Agent That Updates Itself During Operation

Hexo Labs has released SIA to the public — a self-improving agent system based on the gpt-oss-120b model. The project is published under the MIT license, allowing developers to freely use and modify the code.

How the System Works

SIA implements an unusual approach to self-improvement. After each run, a special Feedback-Agent analyzes the task execution trajectory and identifies where the agent made errors or suboptimal decisions. Based on this analysis, the system decides which of two improvement mechanisms to use.

Two Improvement Mechanisms

Hexo Labs implemented two independent methods by which the agent can improve its own performance:

• Scaffold rewriting — changing the agent's instructions, logic, and working structure
• LoRA weight updates — fine-tuning neural network weights without full retraining
• Combined approach — using both mechanisms simultaneously for maximum results

Key finding: when the agent can change both instructions and weights simultaneously, it improves much faster than when using only one method.

Results on Real-World Tasks

The team tested SIA on three complex benchmarks. On LawBench (where the system must solve legal tasks), the combined approach significantly outperformed the results. The same occurred on TriMul — a set of GPU kernels requiring optimization. The third test: denoising single-cell RNA sequencing data (scRNA-seq) — showed that the system is capable of working with bioinformatics.

"This opens the path to agents that not only perform tasks but learn and improve in the process," the

Hexo Labs publication states.

What This Means

The release of SIA to the public is a signal that self-improving agents are transitioning from the category of experiments to a development tool. For researchers, this is an opportunity to test new ideas faster. For companies, this means that soon they will be able to deploy systems that not only perform tasks but become more efficient with each experience.