Most AI agents stop improving once a human stops tuning them. The model stays fixed and the surrounding scaffold never evolves, which limits performance on new or shifting tasks. Teams waste time manually rewriting prompts, adjusting tools, or retraining models in separate cycles, and gains plateau quickly.
Hexo Labs introduced SIA (Self‑Improving AI) to solve this exact problem. SIA treats an agent as two coupled parts: the harness (scaffold) that contains the system prompt, tool dispatch, retry logic and answer extraction, and the model weights themselves. A three‑agent loop drives continuous improvement. A Meta‑Agent creates an initial harness from a task description. A Task‑Specific Agent runs the current harness and logs every step. A Feedback‑Agent reads the full trajectory and decides whether to rewrite the harness while keeping weights fixed, or to update the weights with a low‑rank adapter while keeping the harness unchanged. This decision is made after each run, allowing the system to switch between software‑engineering improvements and internal knowledge acquisition as needed.
The framework was tested on three unrelated benchmarks: legal charge classification, CUDA kernel speed‑up for a biology model, and single‑cell RNA denoising. In every case, combining harness edits with weight updates surpassed harness‑only improvements and exceeded prior state‑of‑the‑art results. The harness gains came from better prompts, tighter parsers and smarter retry policies. The weight gains surfaced domain‑specific knowledge that no prompt could encode, such as efficient GPU kernel patterns or integer‑rounding steps for count data.
SIA is released as open source under an MIT license. It runs on the gpt‑oss‑120b base model with LoRA rank 32, uses Claude Sonnet 4.6 for the meta and feedback agents, and can be launched with a simple command line interface. Four bundled tasks ship with the package, letting teams start self‑improving experiments immediately without manual scaffold tuning or separate weight‑training pipelines.
By editing both the agent’s scaffold and its model weights inside a single loop, SIA removes the need for constant human intervention and delivers continual performance gains across diverse problems.
#AI #Product #MachineLearning #OpenSource #SelfImprovingAI #Tech