essay
GitHub Pulse: Week of March 1, 2026
Agents learned to work in teams, SpacetimeDB collapsed the stack, and WiFi replaced cameras for human sensing.
The agent stack is crystallizing fast. ByteDance, HuggingFace, and Anthropic all shipped composable skill systems this week—agents that call other agents with standardized interfaces. Meanwhile, SpacetimeDB went viral by eliminating the entire app server layer, and WiFi DensePose proved you don't need cameras to see people.
🤖 AI Agents & Assistants (7 projects)
The dominant theme at 47% of trending. The shift: agents are no longer monoliths. They’re composable skill systems.
Three major players shipped agent skill frameworks in the same week: ByteDance’s DeerFlow (orchestration harness), HuggingFace’s Skills (ML workflow modules), and the Superpowers framework (software engineering methodology). The convergence isn’t coincidental—the industry is standardizing how agents delegate subtasks to specialized sub-agents, and the winning pattern is emerging: skills as atomic units, orchestrators as conductors.
obra/superpowers
⚡ VIRAL | 68.1k ★ (+1.3k/day)
An agentic skills framework & software development methodology that works.
Best for: Developers who want their AI coding agents to follow structured software engineering methodology instead of ad-hoc generation Try it: Available as a plugin for Claude Code, Cursor, Codex, and OpenCode Stack: Composable skills framework (20+ modular skills)
| Metric | Value | Read |
|---|---|---|
| Velocity | 9.2x normal | 🔴 Viral spike |
| Fork ratio | 9.4% | 🟡 Moderate |
| Maintenance | Active | 🟢 Recent commits |
Why it matters: Structures agent development into seven phases (brainstorm → plan → execute → test → review → finalize) with modular skills for TDD, debugging, code review, and parallel subagent coordination. This is the shift from “write me code” to “execute software engineering.”
Signal: ⚡ VIRAL The 68k stars reflect genuine adoption across multiple agent platforms. This is becoming the default methodology layer for agent-assisted development.
bytedance/deer-flow
🟢 STRONG | 🏢 ByteDance | 23.5k ★ (+396/day)
An open-source SuperAgent harness that researches, codes, and creates—handling tasks that take minutes to hours.
Best for: Teams building autonomous multi-step workflows (research, report generation, coding) that need sandboxed execution Try it: Docker-based setup · Medium complexity Stack: Python · LangChain · LangGraph · Docker sandboxes
| Metric | Value | Read |
|---|---|---|
| Velocity | 5.0x normal | 🟡 High growth |
| Fork ratio | 11.9% | 🟢 High usage |
| Maintenance | Active | 🟢 Recent commits |
Why it matters: ByteDance’s bet on open agent infrastructure. DeerFlow orchestrates parallel sub-agents inside sandboxed Docker containers with persistent cross-session memory and aggressive context compression. This bridges the gap between simple tool-using chatbots and real execution environments capable of hour-long autonomous workflows.
Signal: 🟢 STRONG MIT licensed, 12% fork ratio proves real usage beyond curiosity. The sandboxing story is critical—production agents need isolation.
shareAI-lab/learn-claude-code
🟢 STRONG | 20.4k ★ (+373/day)
Bash is all you need — A nano Claude Code–like agent, built from 0 to 1
Best for: Developers wanting to understand how agentic coding tools work by building one from scratch Stack: TypeScript · Python · Anthropic SDK
| Metric | Value | Read |
|---|---|---|
| Velocity | 4.5x normal | 🟡 High growth |
| Fork ratio | 19.9% | 🟢 Very high usage |
| Maintenance | Active | 🟢 Recent commits |
Why it matters: Reconstructs Claude Code’s architecture from scratch in 12 progressive sessions: bare “one loop + bash” core → planning → subagents → context compression → multi-agent team coordination. The 20% fork ratio is remarkable—this is educational infrastructure, not just a star magnet. Timing is perfect as Claude Code’s adoption explodes.
Signal: 🟢 STRONG The best way to understand agents is to build one. This is that path.
muratcankoylan/Agent-Skills-for-Context-Engineering
🟡 WATCH | 🧪 Experimental | 13.1k ★ (+624/day)
A comprehensive collection of Agent Skills for context engineering, multi-agent architectures, and production agent systems.
Best for: Teams debugging agent failures related to context window management, attention scarcity, and information poisoning Stack: Python · Curated reference with examples
| Metric | Value | Read |
|---|---|---|
| Velocity | 3.4x normal | 🟡 High growth |
| Fork ratio | 7.7% | 🟡 Moderate |
| Maintenance | Active | 🟢 Recent commits |
Signal: 🟡 WATCH Context engineering is emerging as the defining skill gap between teams shipping reliable agents and teams that aren’t. This is the most comprehensive free reference—organized into four tiers from foundational mechanics to cognitive modeling. Reference material, not a tool, but highly useful reference material.
alibaba/OpenSandbox
🟡 WATCH | 🏢 Alibaba | 4.2k ★ (+287/day)
General-purpose sandbox platform for AI applications: Coding Agents, GUI Agents, Agent Evaluation, RL Training.
Best for: Platform teams needing unified sandbox infrastructure across multiple agent use cases Stack: Python · Docker · Kubernetes · Multi-language SDKs (Python, Java, JS, C#)
| Metric | Value | Read |
|---|---|---|
| Velocity | 5.1x normal | 🟡 High growth |
| Fork ratio | 7.0% | 🟡 Moderate |
| Maintenance | Active | 🟢 Recent commits |
Signal: 🟡 WATCH Alibaba’s answer to agent sandboxing. Built-in code interpreters, filesystem access, per-sandbox egress policies. The unified API layer means teams don’t wire up separate isolation for each agent. Watch for enterprise adoption where code execution security is non-negotiable.
🛠️ Developer Tools (3 projects)
WiFi sensing, code knowledge graphs, and Stanford coursework—three very different signals.
ruvnet/wifi-densepose
⚡ VIRAL | 21.7k ★ (+1.1k/day)
WiFi DensePose turns commodity WiFi signals into real-time human pose estimation, vital sign monitoring, and presence detection—all without a single pixel of video.
Best for: Eldercare monitoring, smart home presence detection, disaster rescue, and any scenario where camera-based sensing raises privacy concerns Try it: Docker deployment in ~30 seconds · ESP32-S3 hardware Stack: Rust · PyTorch · NumPy · SciPy
| Metric | Value | Read |
|---|---|---|
| Velocity | 13.9x normal | 🔴 Viral spike |
| Fork ratio | 11.8% | 🟢 High usage |
| Maintenance | Active | 🟢 Recent commits |
Why it matters: Detects 17 body keypoints, breathing rate, and heart rate through walls up to 5 meters away using cheap ESP32-S3 hardware. Processes 54,000 frames/sec. The privacy-first angle is the killer feature—no cameras, no wearables, no labeled training data required. This sidesteps camera-based privacy debates entirely.
Signal: ⚡ VIRAL The Rust implementation and cheap hardware story make this genuinely deployable, not just a research demo.
abhigyanpatwari/GitNexus
⚡ VIRAL | 8.2k ★ (+827/day)
The Zero-Server Code Intelligence Engine — drop in a GitHub repo and get an interactive knowledge graph with a built-in Graph RAG Agent.
Best for: Developers who want AI coding agents to understand architectural relationships (call chains, blast radius, symbol context) before making changes Stack: TypeScript · Tree-sitter (12+ languages) · MCP compatible
| Metric | Value | Read |
|---|---|---|
| Velocity | 21.3x normal | 🔴 Viral spike |
| Fork ratio | 12.1% | 🟢 High usage |
| Maintenance | Active | 🟢 Recent commits |
Why it matters: Current AI coding agents lack graph-level architectural awareness and frequently break call chains. GitNexus precomputes that structure using Tree-sitter, then serves it via MCP to editors like Claude Code, Cursor, and Windsurf. Runs fully in-browser or as CLI—no server needed.
Signal: ⚡ VIRAL Fills a real gap. The 21x velocity spike suggests developers are actively frustrated with agents that don’t understand code structure.
mihail911/modern-software-dev-assignments
🟡 WATCH | 2.8k ★ (+99/day)
Assignments for CS146S: The Modern Software Dev (Stanford University Fall 2025)
Best for: Students and self-learners interested in Stanford’s take on modern software engineering practices Stack: Python
| Metric | Value | Read |
|---|---|---|
| Velocity | 7.3x normal | 🟡 High growth |
| Fork ratio | 22.1% | 🟢 Very high usage |
| Maintenance | Slow | 🔴 Course complete |
Signal: 🟡 WATCH The 22% fork ratio is the highest this week—people are actually doing the assignments. Educational content, not a tool.
⚙️ LLM Infrastructure (3 projects)
A self-learning vector DB, a skills standard for ML workflows, and edge speech recognition.
huggingface/skills
⚡ VIRAL | 🏢 HuggingFace | 7.9k ★ (+819/day)
Standardized, reusable skill definitions for AI agents to execute ML workflows.
Best for: AI agents that need to reliably execute HuggingFace workflows—dataset creation, model training, evaluation, Gradio UI generation, Hub publishing Stack: Python · Compatible with Claude Code, Codex, Gemini CLI, Cursor
| Metric | Value | Read |
|---|---|---|
| Velocity | 10.2x normal | 🔴 Viral spike |
| Fork ratio | 5.9% | 🟡 Moderate |
| Maintenance | Active | 🟢 Recent commits |
Why it matters: 10+ reusable skill definitions that any AI agent can invoke by name (“use the model trainer skill”). Skills package step-by-step instructions, helper scripts, and guardrails into modules. This abstracts away HuggingFace API complexity so agents can execute full ML workflows without memorizing API patterns.
Signal: ⚡ VIRAL HuggingFace betting that “skills” become the standard interface between agents and ML infrastructure. If this pattern wins, every ML platform will need a skills layer.
ruvnet/ruvector
⚡ VIRAL | 2.5k ★ (+253/day)
High Performance, Real-Time, Self-Learning, Vector Graph Neural Network and Database built in Rust.
Best for: Teams wanting a vector database that improves search quality automatically from query patterns, without manual reindexing Stack: Rust · 230+ SQL functions · Cypher query support
| Metric | Value | Read |
|---|---|---|
| Velocity | 10.4x normal | 🔴 Viral spike |
| Fork ratio | 10.0% | 🟡 Moderate |
| Maintenance | Active | 🟢 Recent commits |
Signal: ⚡ VIRAL The “self-learning” angle is interesting—an embedded GNN that continuously learns from query patterns to improve rankings. Drop-in pgvector replacement with portable cognitive container files that boot in 125ms. Early stage but worth tracking.
moonshine-ai/moonshine
⚡ VIRAL | 6.7k ★ (+307/day)
Fast and accurate automatic speech recognition (ASR) for edge devices
Best for: Any application needing fully offline, privacy-first voice interfaces across mobile, desktop, and embedded devices Stack: C · Python · iOS · Android bindings · 8 language models
| Metric | Value | Read |
|---|---|---|
| Velocity | 23.5x normal | 🔴 Viral spike |
| Fork ratio | 4.6% | 🔴 Low |
| Maintenance | Active | 🟢 Recent commits |
Signal: ⚡ VIRAL Cache-aware streaming architecture achieves ~5x lower latency than Whisper while hitting lower word-error rates than Whisper Large V3 with far fewer parameters. Covers English, Spanish, Mandarin, Japanese, Korean, Vietnamese, Ukrainian, Arabic. The edge-first, fully offline approach fills a real gap in the ASR space.
📦 Other Notable Projects (2 projects)
clockworklabs/SpacetimeDB
⚡ VIRAL | 22.0k ★ (+362/day)
Collapse your entire backend into a single binary. Clients connect directly to the database; application logic runs inside it.
Best for: Real-time applications (games, collaborative tools) where the traditional client → app server → database stack adds unnecessary latency and ops complexity Stack: Rust · Write-ahead log durability · In-memory state
| Metric | Value | Read |
|---|---|---|
| Velocity | 16.3x normal | 🔴 Viral spike |
| Fork ratio | 3.7% | 🔴 Low |
| Maintenance | Active | 🟢 Recent commits |
Why it matters: Eliminates the entire middle tier. The team proves production viability by running BitCraft Online—a full MMORPG—entirely on a SpacetimeDB module with no separate servers. No Kubernetes, no Docker orchestration, no microservices. For real-time apps, this is a provocative simplification of the stack.
Signal: ⚡ VIRAL Low fork ratio suggests more curiosity than adoption so far, but the architectural thesis is compelling. Watch for adoption in gaming and collaborative tools.
moeru-ai/airi
🟡 WATCH | 21.4k ★ (+344/day)
Self-hosted AI virtual companion with real-time voice chat, avatar rendering, and game integration.
Best for: Users wanting a fully customizable, self-hosted AI companion with data sovereignty Stack: TypeScript · WebGPU/WASM · NVIDIA CUDA · Apple Metal · 20+ LLM backends
| Metric | Value | Read |
|---|---|---|
| Velocity | 7.3x normal | 🟡 High growth |
| Fork ratio | 9.4% | 🟡 Moderate |
| Maintenance | Active | 🟢 Recent commits |
Signal: 🟡 WATCH The primary open alternative to closed AI companion platforms. Features real-time voice chat, VRM/Live2D avatars, Minecraft/Factorio gameplay, and Discord/Telegram integration. Niche but significant for the self-hosted AI community.
🔍 Pattern Watch
| Pattern | Signal |
|---|---|
| Agent skills as a standard | Three independent skill frameworks (Superpowers, HuggingFace Skills, DeerFlow) trending simultaneously. “Agent calls agent via skill interface” is becoming the pattern. |
| Context engineering matures | learn-claude-code and Agent-Skills-for-Context-Engineering trending together. The industry is moving from “prompt engineering” to “context engineering” as the core competency. |
| Edge-first renaissance | WiFi DensePose and Moonshine both bet on local-first, privacy-first architectures. Edge computing isn’t dead—it’s being revived by privacy demands. |
| Stack simplification | SpacetimeDB eliminates the app server; GitNexus runs in-browser; WiFi DensePose deploys via Docker in 30 seconds. The best new tools delete infrastructure. |
| Claude Code catalyst | anthropics/claude-code (72.8k ★), learn-claude-code (20.4k ★), and GitNexus (MCP-native) all trending. Claude Code’s ecosystem is generating its own gravity. |
🚫 Noise Filter
Projects that spiked but show weaker fundamentals:
| Project | Issue |
|---|---|
| x1xhlol/system-prompts-and-models-of-ai-tools | 127k stars but it’s a collection of leaked system prompts, not a tool. No license, ethically gray. Reference material at best. |
| anthropics/claude-code | Trending for good reason (Copilot integration, Opus 4.6) but already well-known—not a discovery this week. |
📈 Tech Stack This Week
Language split:
- Python: 36% (agents, ML infra)
- TypeScript: 21% (dev tools, platforms)
- Rust: 21% (databases, performance-critical systems)
- C: 7% (edge ASR)
Emerging patterns:
- MCP compatibility is now table stakes for dev tools (GitNexus, OpenSandbox)
- Multi-agent orchestration frameworks all converge on LangChain/LangGraph or custom skill systems
- Rust is the default for new infrastructure (SpacetimeDB, RuVector, WiFi DensePose)
What This Means
Three shifts are happening simultaneously:
-
Agents are becoming composable. The monolithic “AI assistant” is giving way to orchestrators that delegate to specialized skill modules. DeerFlow, Superpowers, and HuggingFace Skills all converged on this pattern independently—strong signal that it’s correct.
-
The best new tools delete infrastructure. SpacetimeDB eliminates the app server. GitNexus eliminates the code analysis server. WiFi DensePose eliminates cameras. The winning pitch in 2026 isn’t “we added a feature”—it’s “we removed a layer.”
-
Claude Code is becoming an ecosystem. With 72.8k stars, a Copilot integration, and educational repos teaching how to build clones, Claude Code has crossed from “tool” to “platform.” The MCP-native projects trending alongside it (GitNexus, OpenSandbox) confirm the gravitational pull.
The question for next week: Does the “agent skills” pattern consolidate around one standard, or do we get three competing ecosystems?
GitHub Pulse is a weekly analysis of what developers are building. Signal, not noise.