Today's GitHub trending paints a consistent picture: the conversation around AI coding agents has shifted from "will they work?" to "how do we make them work like senior engineers?" The repos drawing the most attention are focused on discipline, structure, and purpose-built infrastructure — not raw capability.
Agent Skills: Encoding Senior-Engineer Habits Into AI Coding Workflows
Addy Osmani of the Google Chrome team released addyosmani/agent-skills, a collection of production-grade workflows packaged as slash commands for AI coding agents, pulling in over 3,000 stars today. The project addresses a familiar frustration: agents tend to take the shortest path to a working result, skipping spec writing, security reviews, and proper test coverage along the way. Agent Skills counters this with seven lifecycle commands (/spec, /plan, /build, /test, /review, /code-simplify, /ship) backed by 20 skill definitions drawn from engineering principles like Hyrum's Law and the Beyonce Rule, plus three specialized reviewer personas — code reviewer, test engineer, and security auditor — for targeted inspection passes.
github.com/addyosmani/agent-skills
PageIndex: Vectorless RAG That Reasons Through Documents
VectifyAI/PageIndex picked up nearly 1,000 stars today with a bet against vector databases for document retrieval. Instead of chunking text and ranking by embedding similarity, PageIndex converts PDFs into hierarchical tree structures that mirror a document's natural section layout, then uses an LLM to reason through that tree to locate relevant content. A financial analysis system built on PageIndex scored 98.7% on the FinanceBench benchmark, significantly outperforming traditional RAG on domain-specific professional documents — where the project's core argument lands: similarity isn't relevance, and complex documents require reasoning to navigate correctly.
github.com/VectifyAI/PageIndex
InsForge: An Open-Source Backend Built for Coding Agents
InsForge/InsForge crossed 460 new stars today with a backend platform that rethinks the usual developer-tool assumption. Rather than treating AI agents as external API consumers, InsForge exposes backend context — schemas, logs, deployed resources — directly to agents and lets them configure authentication, storage, database, and compute infrastructure autonomously via an MCP server or CLI. The platform ships with PostgreSQL, S3-compatible storage, an OpenAI-compatible model gateway, edge functions, and site deployment built in, with self-hosting options via Railway, Zeabur, and Sealos.
DFlash: Faster LLM Inference via Block Diffusion Speculative Decoding
z-lab/dflash gained 671 stars today with a lightweight draft model that accelerates LLM inference using block diffusion — generating multiple candidate tokens in parallel rather than sequentially. A smaller DFlash model proposes token batches; the target LLM verifies and accepts them, cutting wall-clock latency without changing the weights being served. It integrates with vLLM, SGLang, Transformers, and MLX (Apple Silicon), and ships pre-trained draft models covering Qwen3, LLaMA-3.1, and Gemma-4 architectures.
Local Deep Research: Privacy-First Agentic Research at 95% Accuracy
LearningCircuit/local-deep-research added 559 stars today with an agentic research tool designed to run entirely on your own hardware. It combines 10+ search engine integrations — including arXiv, PubMed, Semantic Scholar, and Wikipedia — with local LLM backends (Ollama, LM Studio, llama.cpp) and AES-256 encrypted per-user databases. Cloud models are supported too, and when configured with GPT-4.1-mini and SearXNG it hits roughly 95% accuracy on the SimpleQA benchmark, with a community leaderboard on Hugging Face for comparing setups across different models.