A native Rust re-implementation of the SIA self-improving-agent framework

Abstract

SIA (Self-Improving AI with Harness & Weight Updates) is a framework in which a Meta-Agent writes and improves a Target-Agent, the Target-Agent executes a task, and a Feedback-Agent analyzes the resulting trajectory to propose the next improvement (Hebbar et al., arXiv:2605.27276). This paper is a system and experience report on sia_rust, a native Rust re-implementation of the SIA framework plus a set of safety, observability, and scheduling extensions.

We describe a faithful, byte-parity port of the deterministic SIA core; a native multi-provider LLM agent-runner layer built on injectable transports and tested offline; and a set of extensions scoped honestly as engineering primitives — a capability allow-list and threat model, a native verifier trait, per-generation telemetry, a dashboard, an adaptive scheduler, and a tested pure-CPU LoRA reference. The Rust core runs roughly 5.8× faster by geometric mean across nine operations. We are explicit about what is not yet evaluated and what remains future work.