AI Agents Go Vertical: 4 GitHub Projects Signal Shift Beyond Coding

archify: Verifiable Architecture Diagrams from Code

tt-a1i/archify surged from 8,800 to 26,626 stars in one week — a 3× jump. It generates architecture diagrams directly from a codebase, with each node traceable to specific files and Git commits. Identical inputs produce identical outputs, enabling architecture snapshots that can be diffed for PR reviews.

Last month cathrynlavery/diagram-design focused on brand-adapted aesthetics — reading a site's color palette in 60 seconds to make diagrams look good. archify targets verifiability: nodes link to source locations so diagrams never drift from code. The 3× star growth suggests "accurate" beats "pretty" for core engineering pain points; architecture documentation rot affects every team larger than 10 people.

gods-eye-view: 3D OSINT Simulator with Voice Agent

bilawalsidhu/gods-eye-view launched August 24 and hit 17K stars in 11 days (10,679 this week). It is not an AI agent toolchain but an open-source spatial-intelligence simulator running in the browser, aggregating 13 real-time public-signal layers — aircraft, ships, satellites, earthquakes, public cameras, radio, rocket launches — on a 3D globe.

Author Bilawal Sidhu is a former Google Maps PM; his YouTube series has 5M+ views. Eleven of 13 layers need no API key; npm install && npm run dev runs it locally. The project entered the radar because of a design choice: integration of OpenAI Realtime API voice agent with 28 voice tools, letting users "fly planes by mouth." Click a flight, ask "What is this?", and the agent answers using that plane's live telemetry. AI agent interaction is evolving from 2D chat boxes to multimodal 3D immersive dialogue. July's koala73/worldmonitor built a 2D global-intelligence dashboard; gods-eye-view's 3D + voice agent is more radical. OSINT never lacked signals — it lacked interfaces.

scientific-agent-skills: Skills Enter the Lab

K-Dense-AI/scientific-agent-skills added 7,558 stars this week (42,337 total). It provides 165 verified skills covering biology, chemistry, medicine, and drug discovery, backed by 100+ scientific databases, compatible with Cursor, Claude Code, and Codex.

Skill standardization has been expanding outward: from general coding (mattpocock/skills) to security (mukul975/Anthropic-Cybersecurity-Skills) to reverse engineering (zhaoxuya520/reverse-skill), and now to research. The same week Imbad0202/academic-research-skills also trended — end-to-end academic research skills — confirming research skills are not a one-off. The project claims 190,000+ scientists using it (unverified), but even halved, researcher adoption of agent skills exceeds expectations.

OpenMAIC: Multi-Agent Classroom

THU-MAIC/OpenMAIC from Tsinghua University's MAIC team gained 9,426 stars this week. Input a topic; minutes later receive a full lesson: AI teacher voice lecture, synchronized whiteboard drawing, AI students raising hands and asking questions, real-time quiz grading.

Under the hood, LangGraph orchestrates multiple agents — not one LLM pretending to be a teacher, but distinct agents for lecturer, teaching assistant, and virtual students of varying proficiency. Backed by a JCST'26 paper. Twelve UI languages including Traditional Chinese. DeepSeek Harness integration already exists. July's HKUDS/DeepTutor offered personalized tutoring via a single agent; OpenMAIC's multi-agent classroom mirrors real teaching more closely — education is multi-role interaction, not Q&A.

obscura: Rust-Native Headless Browser for Agents

h4ckf0r0day/obscura (22K+ stars) rewrites a headless browser from scratch in Rust. Claimed 30 MB memory and 85 ms page load versus Headless Chrome's 200 MB+ and ~500 ms (unverified). The technical path is distinctive: no Chromium, single Rust binary. Full Chrome DevTools Protocol implementation means Puppeteer/Playwright scripts theoretically drop in unchanged. Built-in MCP server connects directly to Claude Code and Cursor. DOM-to-Markdown conversion aids RAG scenarios — agents skip extra HTML parsing.

May's CloakHQ/CloakBrowser pursued source-level stealth Chromium; obscura abandons Chromium entirely. The agent browser layer is splitting into "Chrome wrappers" and "native rewrites."

Four Verticals, One Signal

Architecture : archify — verifiable architecture diagrams

Research : scientific-agent-skills — 165 scientific skills

Education : OpenMAIC — multi-agent classroom

Browser : obscura — Rust native headless browser

The general coding agent lane is crowded. Recent weeks moved from "building capabilities" to "behavior governance" to "runtime frameworks" — the infrastructure is built. This week's signal says the next wave is vertical scenarios. Skills are unfolding layer by layer per domain. Agent interaction is expanding from 2D chat to 3D immersive dialogue. The browser layer is diverging from Chrome wrappers to native rewrites. If you're hunting for the next agent opportunity, stop staring at general coding.