Summary

Gliding Horse (LiuMa) is an open-source AI Agent operating system written in Rust, built on the philosophy of 'trust AI, but restrain it with code.' Instead of relying on prompt engineering or model self-discipline, it enforces runtime behavioral constraints through several engineered mechanisms. Its core scheduler hardcodes the PDCA cycle into five distinct Agent roles (Scheduler, Planner, Doer, Checker, Arbiter), with dynamically assigned complexity levels L0-L6 ranging from instant answers to month-long engineering projects. Memory uses a CPU-cache-inspired four-layer hierarchy (Sled/Qdrant, context windows, Oxigraph RDF graph, SPARQL projection) with an MESI-style consistency protocol for multi-Agent shared state. All data exchange uses JSON-LD 1.1 semantic bus, and tool safety is enforced by a syscall gate plus ToolGuard post-call scanning, replacing soft prompt constraints with hard permission matrices. Quality gates use W3C SHACL contracts and 5W2H dimensional audits. The federated architecture pairs a Go-based center (Temporal workflows) with a Rust edge daemon and a VS Code plugin, while a self-evolving skill graph accumulates knowledge over time. Suited to long-horizon software engineering, multi-Agent collaboration, and compliance-heavy enterprise scenarios; overkill for simple queries or low-latency API calls.