A training-free multi-agent harness lifts open-weight Qwen3.8-27B from 66.8% to 92.4% pass@1 on 100

A training-free multi-agent harness (GVS5H) lifts open-weight Qwen3.8-27B from 66.8% to 92.4% pass@1 on 100 hard LiveCodeBench problems by coordinating five model instances through a shared-filesystem ledger — no fine-tuning, pure orchestration effect. A larger orchestrated variant reportedly reaches 93.0% against Claude Fable 5's 90.4%, at 9% of the cost, though the authors themselves disclose the comparison baseline wasn't run under equivalent conditions.

Cloudflare's security-audit-skill leads GitHub Trending for a third consecutive day (45,122 total stars, up from 39,960 on 10-23), and NVIDIA's OpenShell holds #2 for a second straight day on the board (19,611 total) — the 'single-repo skill seeds a fleet-wide internal system' pattern now has a measured multi-day growth curve, not just a launch-day spike.

Claude and ChatGPT are both absent from Google Play's top-30 chart on the same day for the first time in this project's nine-day tracking window — Claude's absence has run since 10-17, but ChatGPT had held a stable top-10 slot throughout until today, leaving Google Gemini as the only frontier-lab AI app left in the capture.

A new arXiv paper formalizes the exact skill-granularity routing problem this project's own research has flagged qualitatively since August: once an agent's skill library outgrows its context budget, deciding what to load and at what granularity becomes the bottleneck. The paper adds a 'macro-to-micro recovery' mechanism that falls back to raw tool calls when a packaged skill fails mid-task.

Block (Jack Dorsey's company) shipped Dessa, a full-duplex voice model that runs locally on a MacBook M4 Max, trained entirely on a single NVIDIA GPU rack. Unlike competitors that use voice tags to steer emotion, Block says emotion is built into the model itself — no public access yet, pending safety review.

Dan Luu's essay arguing AI coding tools justify far more ambitious 'just for fun' software projects trends simultaneously on Hacker News (#2, 304 points) and Lobsters (#6) — a rare same-day cross-platform double, with a dissenting appendix arguing most people should actually build less ambitiously given the lack of clear pass/fail criteria in most real work.

Qwen3.8-27B continues steady download growth on Hugging Face (7,851,161 downloads, up from 7,710,011 the day before) — the same model powering today's multi-agent orchestration benchmark, independently confirming real developer interest beyond one repo's framing.

Two design skills this project already runs (hallmark, for anti-AI-slop design; impeccable, for AI-driven design judgment) are independently trending on GitHub the same day — a coincidental but concrete sign that the design-skill category is now part of the broader agent-tooling wave dominating GitHub Trending (roughly 8 of today's top 10 repos are agent/harness/skill tools).