Hexo Labs Unveils SIA: The Self-Improving Agent That Rewrites Its Own Code and Model Weights

Once a human stops fine-tuning them, most AI agents stop getting better. The underlying model stays the same, and so does the infrastructure built around it. Hexo Labs aims to evolve both simultaneously. This week, the company open-sourced SIA (Self-Improving AI) under the MIT license.

The central — and deliberately narrow — claim behind SIA is this: within a single self-improvement loop, it can modify both the agent’s surrounding infrastructure and the model’s internal weights.

What is SIA (Self-Improving AI)

SIA divides a task-specific agent into two layers. The first layer is the harness (sometimes called the scaffold), which includes the system prompt, tool-routing logic, retry rules, and answer-parsing code. The second layer is the model’s weights themselves.

Three large-language-model components power the loop. A Meta-Agent generates an initial scaffold from a task description and any available reference code. A Task-Specific Agent then runs the job while logging every step of the process. After completion, a Feedback-Agent reviews that full trajectory and determines what adjustments need to be made.

This decision-making step is the core innovation. Following each run, the Feedback-Agent chooses one of two paths: it can revise the scaffold while keeping the model weights frozen, or it can apply a weight update while leaving the scaffold unchanged.

The base model is openai/gpt-oss-120b. For weight updates, SIA uses LoRA — a low-rank adapter — set at rank 32. Both the Meta-Agent and the Feedback-Agent run on Claude Sonnet 4.6. All training is executed on H100 GPUs via Modal, the team’s reinforcement-learning platform.

The team defines two operating modes: SIA-H and SIA-W+H. SIA-H applies harness changes only. SIA-W+H layers weight updates on top of harness edits.

The Benchmark Case

The team evaluated SIA across three deliberately diverse domains. A consistent pattern emerged: weight updates consistently delivered gains beyond what scaffold modifications alone could achieve. “Initial” represents the Meta-Agent’s first scaffold running against the base model, before any feedback loop has fired.

On LawBench, the challenge is a 191-class Chinese criminal-charge classification task. Repeated harness iterations constructed a TF-IDF plus LinearSVC pipeline and leveled off at 50.0% accuracy. Subsequent weight updates through PPO pushed performance up to 70.1% — a 20.1 percentage-point improvement over the harness-only ceiling.

The TriMul task requires writing a custom CUDA kernel for an H100 GPU. The kernel implements a core operation in AlphaFold2’s Evoformer module. Scaffold editing alone achieved a 1.14× speedup over the baseline. Weight updates then slashed runtime from 12,483 µs to 1,017 µs — a 91.9% reduction from the harness-only peak.

The results chart includes an important qualification: the coding assistant Claude Code achieved a 1.50× speedup on TriMul without any help, outpacing SIA-H’s 1.14× result. SIA-W+H still delivered the best overall outcome at 14.02×.

For the denoising task, the agent optimizes MAGIC, a single-cell RNA imputation method. Harness sweeps across hyperparameters settled at an mse_norm of 0.241. After the first weight-update checkpoint, the agent introduced a two-step line that no scaffold iteration had ever generated — it rounded imputed counts to non-negative integers, lifting the score to 0.289.

How the Feedback-Agent Picks Its Move

SIA doesn’t rely on a single fixed RL recipe. The Feedback-Agent chooses a training algorithm based

# install the Claude backend
pip install 'sia-agent[claude]'
export ANTHROPIC_API_KEY="..."

# execute 5 rounds of self-improvement on a bundled task
sia --task lawbench --max_gen 5 --run_id 1