GitHub AI Daily · 2026-10-18 (Sunday · Issue 3)

Picks (4)

1. deepseek-ai/deepseek-harness — a model vendor builds a plug-in agent runtime itself

Repo: https://github.com/deepseek-ai/deepseek-harness ｜ ★240k

DeepSeek's official agent harness is built on "Everything is a Plugin" (the Cordis/DSH plugin system). At 240k stars this is clearly not an experiment: the model vendor is defining the runtime layer rather than leaving it to the community.

2. mattpocock/skills — 156k stars prove "methodology beats model"

Repo: https://github.com/mattpocock/skills ｜ ★156k, still gaining ~1,000/day

A pure-shell .claude skill pack with no framework, no company and no funding. It targets the four places where coding agents actually fail — misreading the task, verbose output, code that does not run, and architecture rot.

3. MedicalHarness — a quarter of a medical agent's score belongs to the harness

Repo: https://github.com/REAL-Lab-NU/MedicalHarness ｜ arXiv:2610.05778

A controlled experiment from Northwestern: hold the model fixed and swap only the harness. The harness and its interaction with the model explain about 25% of the variance in medical task outcomes, and no single harness wins across all models and tasks.

4. The health and nutrition data layer reaches stage two: unified platform plus conversation-as-UI

Repos: https://github.com/akutishevsky/nutrition-mcp ｜ https://github.com/the-momentum/open-wearables

Last issue we saw a cluster of single-purpose health MCP servers; this issue brings two markers of evolution — nutrition-mcp replaces an app with a single conversational surface, and open-wearables consolidates wearable data behind one gateway.

Trends

1. Skills are becoming personalities. mattpocock/skills, agency-agents and ponytail together show agent skills evolving from "tool manuals" into installable engineering personalities and methodologies; a personal .claude directory is a new form of technical asset.
2. Decision models are a new species. vllm-sr Decision 3.0 and Amazon Strands Decider appeared together: a model layer that emits only a choice and a confidence, no tokens. If every routing step of an agent can be done by a 2B model, the cost curve gets rewritten.
3. Medical evaluation enters a three-dimensional era: model × harness (MedicalHarness) × time (HAEnv's longitudinal synthetic patients). Single-turn Q&A benchmarks are becoming obsolete.
4. The health data layer is consolidating. A dozen single-point health MCPs are merging into a few unified platforms (open-wearables, mirobody); nutrition-mcp's "replace the app" narrative points to the next battlefield — migration and data sovereignty.