Key points
- Concept and positioning. Gliding Horse (流马) is an MIT-licensed, industrial-grade AI Agent operating system written in Rust, Go, and TypeScript. The name deliberately echoes Zhuge Liang's wooden ox and gliding horse, framing Agent orchestration as autonomous logistics rather than as a chat framework.
- Generalized PDCA execution model. The PDCA cycle is elevated from a management metaphor into a seven-level computational model (L0 instant query to L6 emergency bypass). A Supervisor Agent dynamically selects PDCA depth based on a 5W2H ontology, replacing rigid templates.
- Five-agent pipeline. A fixed industrial assembly line — Supervisor, Planner, Doer, Checker, Actor — replaces open-ended multi-Agent chat, with each role bounded by explicit responsibilities and a blackboard-backed workflow.
- Five-layer memory with MESI coherence. L1 context window, L2 in-memory RDF, L3 SPARQL projection, and L0 Sled + Qdrant storage adopt CPU cache-coherence semantics (Modified, Exclusive, Shared, Invalid). A prefetch engine driven by Agent intent lowers perceived load latency from about 50 ms to about 2 ms.
- JSON-LD 1.1 as the semantic bus. W3C-standard
@context,@id,@type, and named graphs supply global entity identity, polymorphic discovery, conflict-free parallel writes, and progressive disclosure across five token-economy levels. - Unified knowledge and skill graphs. IRIs act as the bridge between Qdrant vector retrieval and Oxigraph SPARQL traversal, enabling hybrid semantic-plus-symbolic retrieval without duplicate-merge logic.
- Self-evolving skill graph. Roughly 7,500 LOC of skill nodes carry success rate, usage count, and maturity metadata, and grow via experience writeback, a
/learnmechanism, a/reduceextraction step, and an experimental-to-production maturity ladder. - 5W2H as task ontology. Each of the seven dimensions is filled at a specific pipeline stage and audited independently by the Checker, enabling precise rollback (What/Why failure reopens analysis; How/Where failure reopens planning; When/HowMuch failure downgrades or terminates).
- Runtime security via SyscallGate. Tool calls are gated by Ed25519 signature verification, JSON Schema validation, and a permission matrix, then executed inside Docker sandboxes. A 10-trigger perception engine with 60-second anomaly deduplication escalates to human intervention.
- Center-Edge federation. Go + Temporal orchestrate workflows and registries; a Rust + axum edge daemon runs local LLMs and Docker sandboxes; a TypeScript VS Code plugin provides the developer UI. The edge can run offline, removing single points of failure.
- Reported performance. L0 KV reads reach ~1 ms at 1000 ops/sec, L2 graph writes ~2 ms at 500 ops/sec, L3 SPARQL projection ~15 ms at 66 ops/sec, ReAct turns 1-5 s, and idle memory footprint ~200 MB scaling with task size.
- Open design philosophy. The project positions itself against prompt-based safety by asserting that trustworthy autonomy requires engineering-level hard constraints, so the Agent "cannot do wrong even if it wanted to."
- GitHub: https://github.com/doiito/gliding_horse
- Design detail: docs/DESIGN_DETAIL.md
- Design philosophy: docs/CORE_DESIGN_PHILOSOPHY.md
- Protocol: proto/pdca_core.proto
- License: MIT