Agents Are Coming Home: Local Models Meet the Skill Layer

The most interesting AI news this week isn't a frontier benchmark - it's a cluster of signals pointing the same way: the agent stack is moving to the edge, and the harness is becoming the product.

A local workhorse finally arrives

The Edge0-35B-A3B-preview trending on Hugging Face is the headline: a 35B-parameter MoE with only 3B active parameters, still in preview. That ratio is exactly what a 128GB unified-memory machine wants - the full weights fit with room to spare, and 3B active means near-small-model speed with big-model quality. If it holds up out of preview, it beats a 27B-class setup on both quality and speed.

The zero-risk smaller add is Spark-X2.6-4B, a refreshed 4B agent model with stronger coding, agent, and long-context performance, a 3-level controllable thinking budget, and reports of matching ~9B-class models. It's the ideal fast side-model for agentic pipelines.

Nvidia is pushing the same direction. Nemotron 3.5 pairs a Super VL 120B for enterprise workflows with a Bolt 74B-A9B MoE that can be 2-bit-quantized to fit in 24GB - running agent tasks on RTX or DGX Spark. And the demand side is following: open-weight models jumped from 7% to 56% of Vercel gateway tokens in nine months.

The harness is the product

A Hacker News explainer (How LLMs work) puts it bluntly: the real unit of capability is the LLM plus the agent harness, and the model is just the interpreter layer. GitHub trending agrees.

The standout is graphify - it turns any codebase (docs, SQL schemas, configs, even PDFs) into a queryable knowledge graph via deterministic local AST parsing, with no vector store, and ships as a drop-in skill for Claude Code, Cursor, Codex, and Gemini CLI. It's adding more than 1,100 stars a day.

Cloudflare's security-audit-skill runs multi-phase security audits through coding agents and produces independently verified, machine-readable findings - verification loops instead of vibes. And deepseek-harness takes an "everything is a plugin" approach: adaptive memory, self-learning, RAG, with native Claude Code, Codex, and Hermes integration.

The pattern is the point: four of the top-10 velocity repos today are Claude Code/Cursor skills. The center of gravity is shifting from models to the skill and workflow layer around agents.

Why it matters

When a 35B-class model runs on laptop-class hardware, model access stops being the moat. The differentiator becomes the harness: structured retrieval instead of embeddings, verification loops, memory. That's the stack you can actually build on today, on the hardware you already own.

Signal to watch: Edge0-35B-A3B's full release. If it holds the 3B-active ratio at 35B quality, it's the obvious local workhorse - and the skill repos piling onto it will show which harness wins.