Today's standout GitHub repositories make a quiet argument: the limiting factor for AI agents isn't model intelligence, it's grounding. Whether that's physical grounding through real-world sensor data, domain grounding through curated scientific APIs, or procedural grounding through disciplined engineering specs — the most-starred projects this Friday all bet on giving agents something more concrete to work with than a blank context window.
RuView: Spatial Intelligence Without a Single Camera
ruvenet's RuView converts commodity WiFi signals into a full sensing stack: contactless presence detection, real-time 17-point body pose estimation, and vital sign monitoring — breathing rate between 6–30 BPM and heart rate between 40–120 BPM — without a single pixel of video. The system runs entirely on edge hardware, with node costs ranging from $8 (ESP32-S3) to $140 for research-grade configurations, and ships 60 WASM inference modules written primarily in Rust. By analyzing channel state information (CSI) — the way WiFi signals scatter differently when someone breathes versus when they walk versus when a room is empty — RuView tracks occupancy through walls, in complete darkness, and without triggering GDPR video-surveillance obligations. At 56,000+ stars and gaining nearly 1,700 today, it's the week's breakout project.
Scientific Agent Skills: 135 Research Capabilities, Ready to Drop In
K-Dense-AI's scientific-agent-skills packages 135 domain-specific capabilities for AI coding agents — Cursor, Claude Code, Codex, and compatible tools — across bioinformatics, cheminformatics, clinical research, materials science, medical imaging, and financial analysis. Each skill includes curated documentation with examples and provides REST API access to over 78 public databases: PubChem, ChEMBL, UniProt, ClinicalTrials.gov, and more. The project's central bet is that agents fail at complex scientific workflows not because models lack knowledge, but because they lack reliable API schemas and a consistent way to invoke specialized tools — structured skills address all three without requiring the agent to rediscover the same contracts on each run.
github.com/K-Dense-AI/scientific-agent-skills
Spec Kit: GitHub's Framework for Spec-Driven AI Development
GitHub's own spec-kit formalizes a development workflow where detailed specifications don't just guide an AI agent — they drive it. Seven slash commands walk a project from first principles (/speckit.constitution) through requirements definition (/speckit.specify), technical planning (/speckit.plan), ordered task generation (/speckit.tasks), and systematic implementation (/speckit.implement), with a consistency validator (/speckit.analyze) designed to surface underspecified requirements before any code is written. The toolkit is language-agnostic, works with 30+ AI coding agents, and requires only Python and uv to install. The underlying argument echoes what teams have been learning from agentic coding deployments: agents with clear specs outperform agents given more capable models and vague briefs, and the right methodology can substitute for several model generations of capability improvement.