GitHub Open Source Weekly 2026-03-18: Agent Harness Ecosystem Matures, Browser Automation Infrastructure Emerges, BitNet Dominates HN with 370 Points
Data period: 2026-03-11 to 2026-03-18 (Rolling 7 days) Sources: GitHub Trending weekly + monthly, GitHub Search API, HN Algolia
TL;DR: Three major themes defined this week. First, the agent harness ecosystem is clearly maturing — obra/superpowers gained +37,809 stars this month, and Y Combinator president Garry Tan's gstack hit 23,057 stars in its first seven days, signaling that "opinionated agent configurations" are becoming mainstream. Second, a browser automation infrastructure layer is emerging simultaneously, with lightpanda-io/browser (a Zig headless browser) and alibaba/page-agent (147 HN points) representing two distinct architectural approaches. Third, microsoft/BitNet became this week's community discussion champion with 370 HN points and 169 comments, debating whether 1-bit LLMs can actually run on edge devices.
📈 Fastest Growing — Weekly Star Growth Top 15
Source:
github.com/trending?since=weekly🔁 = Also appears in monthly trending (sustained momentum signal)
| # | Project | +Stars/week | Total Stars | Language | Created |
|---|---|---|---|---|---|
| 1 | msitarzewski/agency-agents | +29,160 | 53,404 | Shell | 2025-10 |
| 2 🔁 | 666ghj/MiroFish | +18,725 | 33,778 | Python | 2025-11 |
| 3 🔁 | obra/superpowers | +14,768 | 95,036 | Shell | 2025-10 |
| 4 🔁 | volcengine/OpenViking | +9,306 | 15,600 | Python | 2026-01 |
| 5 🔁 | lightpanda-io/browser | +8,700 | 21,531 | Zig | 2023-02 |
| 6 | alibaba/page-agent | +7,000 | 11,126 | TypeScript | 2025-09 |
| 7 | microsoft/BitNet | +6,410 | 35,474 | Python | 2024-08 |
| 8 | pbakaus/impeccable | +6,259 | 10,332 | JavaScript | 2025-11 |
| 9 🔁 | shareAI-lab/learn-claude-code | +5,582 | 31,896 | TypeScript | 2025-06 |
| 10 | promptfoo/promptfoo | +5,473 | 17,367 | TypeScript | 2023-04 |
| 11 | NousResearch/hermes-agent | +4,991 | 8,695 | Python | 2025-07 |
| 12 | langchain-ai/deepagents | +3,520 | 14,862 | Python | 2025-07 |
| 13 | fishaudio/fish-speech | +2,775 | 28,111 | Python | 2023-10 |
| 14 | langflow-ai/openrag | +2,533 | 3,231 | Python | 2025-07 |
| 15 | vectorize-io/hindsight | +1,996 | 4,799 | Python | 2025-10 |
🆕 Top New Repos — Best New Projects This Week
Source: GitHub Search API (
created:2026-03-11..2026-03-18, sorted by total stars)
| # | Project | Total Stars | Language | Created |
|---|---|---|---|---|
| 1 | garrytan/gstack | 23,057 | TypeScript | 2026-03-11 |
| 2 | THU-MAIC/OpenMAIC | 6,762 | TypeScript | 2026-03-11 |
| 3 | NVIDIA/NemoClaw | 6,153 | TypeScript | 2026-03-15 |
| 4 | aiming-lab/AutoResearchClaw | 5,934 | Python | 2026-03-15 |
| 5 | calesthio/Crucix | 3,849 | JavaScript | 2026-03-14 |
| 6 | webadderall/Recordly | 2,457 | TypeScript | 2026-03-12 |
| 7 | pasky/chrome-cdp-skill | 2,160 | JavaScript | 2026-03-13 |
| 8 | davebcn87/pi-autoresearch | 2,155 | TypeScript | 2026-03-11 |
| 9 | TianyiDataScience/openclaw-control-center | 2,093 | TypeScript | 2026-03-11 |
| 10 | gsd-build/gsd-2 | 1,989 | TypeScript | 2026-03-11 |
| 11 | MoonshotAI/Attention-Residuals | 1,738 | — | 2026-03-15 |
| 12 | jackwener/opencli | 1,677 | TypeScript | 2026-03-14 |
| 13 | novatic14/MANPADS-System-Launcher-and-Rocket | 1,637 | C++ | 2026-03-11 |
| 14 | Narcooo/inkos | 1,618 | TypeScript | 2026-03-12 |
| 15 | uditgoenka/autoresearch | 1,314 | Shell | 2026-03-13 |
Spotlight — Fastest Growing Top 15
📈 #1 — msitarzewski/agency-agents｜A Complete AI Agency in Your Hands
A complete AI agency at your fingertips - From frontend wizards to Reddit community ninjas, from whimsy injectors to reality checkers. Each agent is a specialized expert with personality, processes, and proven deliverables.
+29,160 ★ this week｜53,404 total｜Shell｜MIT
agency-agents has topped the weekly trending chart for two consecutive weeks. Its design philosophy: every AI agent isn't just "a prompt" — it's a specialized role with personality, processes, and defined deliverables, from frontend wizards and Reddit community ninjas to creativity injectors and reality checkers.
Starting from Shell scripts makes it easy to integrate into any CI/CD environment. Two weeks of sustained growth to 53K stars reflects strong developer consensus around the idea of "managing agents like team members."
📈 #2 — 666ghj/MiroFish 🔁｜Swarm Intelligence Engine for Predicting Anything
A Simple and Universal Swarm Intelligence Engine, Predicting Anything.
+18,725 ★ this week｜33,778 total｜Python｜AGPL-3.0
MiroFish is the fastest-growing sustained hit this week, jumping from +8,983 last week to +18,725. It uses swarm intelligence simulation for prediction tasks — financial forecasting, sentiment analysis, social dynamics modeling — with knowledge graphs, multi-agent simulation, and agent memory under the hood.
AGPL-3.0 licensing plus 125 open issues indicates a highly active but rapidly evolving project. Verify commercial use terms before adopting.
📈 #3 — obra/superpowers 🔁｜The Agent Harness Benchmark at 38K Monthly Stars
An agentic skills framework & software development methodology that works.
+14,768 ★ this week｜95,036 total｜Shell｜MIT
superpowers is the highest monthly gainer among all trending repos this week (+37,809 monthly stars), sitting just under 100K stars. It's not just a tool — it's a complete AI-assisted software development methodology that modularizes agent capabilities into composable, extensible, version-controlled skills.
If you're evaluating whether to build your own agent harness, superpowers is the most community-validated reference implementation available. Read it alongside this week's new repo chart topper garrytan/gstack for a fuller picture of where this ecosystem's boundaries lie.
📈 #4 — volcengine/OpenViking 🔁｜ByteDance's Agent Context Database
OpenViking is an open-source context database designed specifically for AI Agents. OpenViking unifies the management of context (memory, resources, and skills) that Agents need through a file system paradigm, enabling hierarchical context delivery and self-evolving.
+9,306 ★ this week｜15,600 total｜Python｜Apache-2.0
OpenViking, from ByteDance's Volcengine, is this week's most representative repo for the "agent infrastructure layer." Its core premise: the ceiling on agent capability isn't the model — it's context management. It unifies memory, resources, and skills in a filesystem paradigm with hierarchical context delivery and self-evolution support.
Also in monthly trending (+12,866 monthly), showing sustained attention. Compared to vectorize-io/hindsight (#15) which focuses on learned agent memory, the two represent different entry points into agent infrastructure.
📈 #5 — lightpanda-io/browser 🔁｜A Headless Browser Built for the AI Era in Zig
Lightpanda: the headless browser designed for AI and automation
+8,700 ★ this week｜21,531 total｜Zig｜AGPL-3.0
Lightpanda is one of two browser automation infrastructure picks this week — a headless browser reimplemented in Zig specifically for AI agents and automation, supporting the Playwright/Puppeteer-compatible CDP protocol. Zig's choice delivers lower memory usage and more predictable performance, particularly appealing for agent scenarios requiring many concurrent browser instances.
Also in monthly trending (+8,945 monthly), indicating sustained development rather than a one-time spike.
📈 #6 — alibaba/page-agent｜In-Page GUI Agent with 147 HN Points
JavaScript in-page GUI agent. Control web interfaces with natural language.
+7,000 ★ this week｜11,126 total｜TypeScript｜MIT
page-agent has the highest-quality community discussion of any Fastest Growing repo this week. Unlike Lightpanda's external approach, page-agent "lives inside" the web page DOM and controls the interface with natural language from within.
The HN Show HN: PageAgent, A GUI agent that lives inside your web app thread got 147 points and 77 comments, centering on: what's the fundamental difference between this and Playwright/Puppeteer? Is the path to MCP toolchain integration clear? The discussion reveals genuine developer debate over "in-process vs. external agent" architecture choices.
📈 #7 — microsoft/BitNet｜1-Bit LLM Inference Framework, 370 HN Points
Official inference framework for 1-bit LLMs
+6,410 ★ this week｜35,474 total｜Python｜MIT
BitNet is this week's absolute community discussion champion: 370 points and 169 comments on HN, far ahead of all other repos. Microsoft Research's 1-bit LLM inference framework represents model weights as 1.58-bit (-1, 0, +1), theoretically enabling sufficiently capable LLMs to run on CPUs or low-power edge devices.
The core HN debate: how significant is the capability loss with 1-bit models? In which tasks is the degradation acceptable? Has edge inference (IoT, mobile, offline environments) actually reached usable quality? This discussion reflects the developer community's sharpest, most realistic pushback on "LLM democratization." If your use case requires edge inference, BitNet is the most current official framework to dig into.
📈 #8 — pbakaus/impeccable｜A Design Language That Makes Your AI Harness Better at Design
The design language that makes your AI harness better at design.
+6,259 ★ this week｜10,332 total｜JavaScript｜Apache-2.0
impeccable takes a rare angle: it's not a design tool, but a "design language specification" for AI agents — giving AI harnesses a clear standard to follow when generating UI, writing CSS, or making design decisions. In a sense, it's the "design track" infrastructure layer for the agent harness ecosystem.
📈 #9 — shareAI-lab/learn-claude-code 🔁｜Build a Nano Claude Code from Scratch
Bash is all you need - A nano claude code–like 「agent harness」, built from 0 to 1
+5,582 ★ this week｜31,896 total｜TypeScript｜MIT
learn-claude-code has appeared in monthly trending for several consecutive weeks (+12,982 monthly), and remains the best learning resource for understanding "how agent harnesses work from the ground up." Starting from a minimal Bash script, it progressively demonstrates Claude Code's core mechanics — ideal for developers who want deep understanding, not just usage.
📈 #10 — promptfoo/promptfoo｜Acquisition Rumors Fuel HN Community Debate
Test your prompts, agents, and RAGs. Red teaming/pentesting/vulnerability scanning for AI.
+5,473 ★ this week｜17,367 total｜TypeScript｜MIT
promptfoo's appearance this week has a notable backstory: an HN thread titled Ask HN: With Promptfoo acquired by OpenAI, what are MCP devs using for testing? explored what developers should use for prompt testing and AI red teaming if promptfoo were acquired by OpenAI. The discussion itself drove a wave of attention to promptfoo — regardless of whether the rumors are true, its position as an integrated LLM eval + red teaming tool is well established.
📈 #11 — NousResearch/hermes-agent｜The Agent That Grows With You
The agent that grows with you
+4,991 ★ this week｜8,695 total｜Python｜MIT
Nous Research's hermes-agent emphasizes "self-growth" — the agent accumulates memory and adjusts behavior through use, rather than starting from zero each session. Topics include openclaw and clawdbot, indicating it's a core OpenClaw ecosystem agent implementation. Reading it alongside OpenViking (#4)'s context database gives a more complete picture of the entire OpenClaw agent ecosystem.
📈 #12 — langchain-ai/deepagents｜LangChain's Official Deep Agent Harness
Agent harness built with LangChain and LangGraph. Equipped with a planning tool, a filesystem backend, and the ability to spawn subagents.
+3,520 ★ this week｜14,862 total｜Python｜MIT
deepagents is LangChain's official heavyweight agent harness, integrating planning, filesystem backend, and subagent spawning as its three core capabilities, positioned for "handling complex tasks requiring minutes to hours." As the LangChain ecosystem's official implementation, it represents the ecosystem's latest design thinking on long-horizon autonomous agents.
📈 #13 — fishaudio/fish-speech｜Sustained Open Source SOTA TTS
SOTA Open Source TTS
+2,775 ★ this week｜28,111 total｜Python｜Custom license
fish-speech has maintained open source SOTA status in TTS for an extended period. This week's renewed attention likely correlates with growing voice agent demand — as more agents need speech output capabilities, high-quality open source TTS becomes increasingly important. Note the NOASSERTION license (non-standard) — confirm commercial use terms before adopting.
📈 #14 — langflow-ai/openrag｜Complete RAG Platform: Langflow + Docling + OpenSearch
OpenRAG is a comprehensive, single package Retrieval-Augmented Generation platform built on Langflow, Docling, and Opensearch.
+2,533 ★ this week｜3,231 total｜Python｜Apache-2.0
openrag's value proposition is "all-in-one RAG" — no need to stitch together components yourself. One package handles the complete pipeline from document parsing (Docling) to vector search (OpenSearch) to flow management (Langflow). Only 3,231 total stars but 2,533 added this week indicates it just entered a breakout phase.
📈 #15 — vectorize-io/hindsight｜Agent Memory That Learns
Hindsight: Agent Memory That Learns
+1,996 ★ this week｜4,799 total｜Python｜MIT
hindsight addresses a concrete problem: how can an agent automatically learn from both successes and failures and update its memory after executing tasks? It's not static RAG — it's a dynamic, evolvable agent memory system. Compared to OpenViking (#4), the two represent different abstraction levels of "memory management" in the agent infrastructure layer.
Spotlight — Top New Repos
🆕 #1 — garrytan/gstack｜Y Combinator President's Claude Code Setup, 23K Stars in Week One
Use Garry Tan's exact Claude Code setup: 10 opinionated tools that serve as CEO, Eng Manager, Release Manager, Doc Engineer, and QA
23,057 total ★｜TypeScript｜MIT｜Created: 2026-03-11
The biggest surprise in this week's new repo chart. gstack is Garry Tan's (Y Combinator President) publicly released Claude Code configuration — 10 opinionated tools each playing distinct roles: CEO, Engineering Manager, Release Manager, Doc Engineer, and QA.
The HN thread Garry Tan's Claude Code Setup got 67 points and 67 comments, centering on: where does this "role-based agent configuration" actually outperform generic configs? Which role setups are genuinely useful vs. performative? 23K stars in seven days reflects developers' intense curiosity about and desire to learn from real-world workflows of well-known practitioners.
🆕 #2 — THU-MAIC/OpenMAIC｜Tsinghua University's Multi-Agent Interactive Classroom
Open Multi-Agent Interactive Classroom — Get an immersive, multi-agent learning experience in just one click
6,762 total ★｜TypeScript｜AGPL-3.0｜Created: 2026-03-11
An open source project from Tsinghua University's MAIC Lab, using multi-agent simulation to create immersive learning environments. The concept: when a student is learning a topic, multiple AI agents playing different roles (Socratic tutor, challenger, encourager) guide deep understanding interactively, rather than through linear instruction.
🆕 #3 — NVIDIA/NemoClaw｜NVIDIA's Official Secure Installation Plugin for OpenClaw
NVIDIA plugin for secure installation of OpenClaw
6,153 total ★｜TypeScript｜Apache-2.0｜Created: 2026-03-15
NVIDIA released NemoClaw this week, providing secure installation mechanisms for OpenClaw with official documentation support (docs.nvidia.com/nemoclaw). This signal matters: when the most important hardware vendor in the GPU space starts actively providing official integrations for an agent ecosystem, it means that ecosystem has crossed into the "vendor endorsement" phase.
🆕 #4 — aiming-lab/AutoResearchClaw｜Fully Autonomous Research Agent from Idea to Paper
Fully autonomous & self-evolving research from idea to paper. Chat an Idea. Get a Paper.
5,934 total ★｜Python｜MIT｜Created: 2026-03-15
AutoResearchClaw continues last week's autoresearch trend, delivering "give it an idea, it produces a paper" fully automated research, with citation verification, multi-agent debate mechanisms, and self-evolution capabilities. Topics include openclaw and metaclaw, indicating it's an OpenClaw ecosystem extension.
🆕 #5 — calesthio/Crucix｜Your Personal Intelligence Terminal
Your personal intelligence agent. Watches the world from multiple data sources and pings you when something changes.
3,849 total ★｜JavaScript｜AGPL-3.0｜Created: 2026-03-14
Crucix makes intelligence monitoring personal — self-host an agent that watches multiple data sources simultaneously and actively notifies you when specific events occur. Following last week's koala73/worldmonitor hitting 32,984 monthly stars, this direction is generating more new entrants this week.
🆕 #6 — webadderall/Recordly｜Free Open Source Screen Studio Alternative
A free, open-source Screen Studio alternative that adds auto-zoom, cursor animations and more to your screen recordings.
2,457 total ★｜TypeScript｜MIT｜Created: 2026-03-12
The only non-agent tool in this week's Top New Repos, and the one closest to everyday developer tooling. Screen Studio's macOS pricing is prohibitive for many, and Recordly provides a complete cross-platform (Windows/macOS/Linux) open source alternative with auto-zoom, cursor animations, and all core features.
🆕 #7 — pasky/chrome-cdp-skill｜Give Your Agent Access to Your Live Chrome Tabs
Give your AI agent access to your live Chrome session — works out of the box, connects to tabs you already have open
2,160 total ★｜JavaScript｜No license｜Created: 2026-03-13
chrome-cdp-skill's positioning creates an interesting contrast with lightpanda-io/browser: the former lets agents connect directly to your already-open Chrome tabs; the latter is a lightweight headless browser built from scratch. One is "leverage your existing browser environment," the other is "give the agent an isolated browser environment" — two routes with distinct use cases. Note: no license currently listed, use with caution commercially.
🆕 #8 — davebcn87/pi-autoresearch｜Autonomous Experiment Loop for pi
Autonomous experiment loop extension for pi
2,155 total ★｜TypeScript｜MIT｜Created: 2026-03-11
Adds autonomous experiment loop capability to pi (an AI assistant platform), enabling it to continuously iterate, verify, and keep or discard hypotheses.
🆕 #9 — TianyiDataScience/openclaw-control-center｜Making OpenClaw Transparent and Controllable
Turn OpenClaw from a black box into a local control center you can see, trust, and control.
2,093 total ★｜TypeScript｜MIT｜Created: 2026-03-11
openclaw-control-center addresses a real pain point: OpenClaw's agent execution is opaque — you don't know what it's doing. This tool converts it into a visual local control center where you can see, trust, and intervene in the entire agent execution flow.
🆕 #10 — gsd-build/gsd-2｜Meta-Prompting System for Agents That Don't Lose the Plot
A powerful meta-prompting, context engineering and spec-driven development system that enables agents to work for long periods of time autonomously without losing track of the big picture
1,989 total ★｜TypeScript｜MIT｜Created: 2026-03-11
gsd-2 focuses on the long-horizon autonomous execution problem: how do you keep an agent on track across hours of complex task work? It uses meta-prompting and spec-driven development to maintain a continuous "big picture" framework. Topics: context-engineering, meta-prompting, spec-driven-development.
🆕 #11 — MoonshotAI/Attention-Residuals｜Moonshot AI Research Repo
1,738 total ★｜No language｜No license｜Created: 2026-03-15
A research repo from Moonshot AI (creators of Kimi), currently without description. The name suggests research related to attention mechanisms and residual connections. The 1,738 stars are primarily driven by Moonshot AI brand recognition.
🆕 #12 — jackwener/opencli｜Turn Any Website Into Your CLI
Make any website your CLI. A powerful, AI-native runtime for seamless browser automation and dynamic web data extraction.
1,677 total ★｜TypeScript｜Apache-2.0｜Created: 2026-03-14
opencli's concept is direct: use AI to translate any website's interaction logic into a CLI interface, useful for batch-processing web operations from the command line.
🆕 #13 — novatic14/MANPADS-System-Launcher-and-Rocket
1,637 total ★｜C++｜No license｜Created: 2026-03-11
A C++ repo without description, with 1,637 stars and 431 forks. The name involves military equipment (MANPADS = Man-Portable Air Defense Systems). Unable to provide an assessment without confirmed technical details.
🆕 #14 — Narcooo/inkos｜AI Agents That Write, Audit, and Revise Novels Autonomously
Autonomous novel writing CLI agent — AI agents write, audit, and revise novels with human review gates
1,618 total ★｜TypeScript｜MIT｜Created: 2026-03-12
inkos brings agent collaboration to creative writing: multiple AI agents divide labor across writing, reviewing, and revising novels, with humans only intervening at key gates. Topics include chinese-novel, indicating a primary focus on Chinese-language creative content.
🆕 #15 — uditgoenka/autoresearch｜Karpathy Autoresearch as a Claude Code Skill
Claude Autoresearch Skill — Autonomous goal-directed iteration for Claude Code. Inspired by Karpathy's autoresearch. Modify → Verify → Keep/Discard → Repeat forever.
1,314 total ★｜Shell｜MIT｜Created: 2026-03-13
Karpathy's autoresearch concept adapted for Claude Code, enabling Claude Code to autonomously execute "modify → verify → keep or discard → repeat forever" goal-directed iteration loops.
Monthly Trend Cross-Reference
Five repos appear in both weekly and monthly trending this week (🔁 markers), sorted by monthly growth:
| Project | Monthly Stars | Notes |
|---|---|---|
| obra/superpowers | +37,809 | Agent harness ecosystem benchmark |
| affaan-m/everything-claude-code | +36,186 | Claude Code optimization bible (monthly leader) |
| 666ghj/MiroFish | +27,746 | Swarm intelligence prediction engine |
| shareAI-lab/learn-claude-code | +12,982 | Claude Code tutorial |
| volcengine/OpenViking | +12,866 | Agent context database |
| lightpanda-io/browser | +8,945 | AI-native headless browser |
Key monthly signal: affaan-m/everything-claude-code (+36,186 monthly) didn't make the weekly Fastest Growing Top 15 but has extremely high monthly cumulative growth — indicating "Claude Code optimization configurations" is a sustained direction with ongoing contributions and adoption, not a temporary spike.
This Week's Trend Analysis
Agent Harness Ecosystem Moving from Experiment to Standardization
This week's clearest signal: agent harness repos are no longer just personal experiments. We're now seeing "renowned practitioners' opinionated configs" (garrytan/gstack), "official LangChain implementations" (langchain-ai/deepagents), and "NVIDIA official integrations" (NemoClaw) appear simultaneously. Three dimensions of validation in the same week means the agent harness ecosystem has crossed the "early adopter toy" threshold.
Browser Automation Entering Infrastructure Phase
Four related repos appeared this week: lightpanda-io/browser (Zig headless browser), alibaba/page-agent (in-page GUI agent, 147 HN points), pasky/chrome-cdp-skill (direct Chrome connection), plus monthly trending alibaba/OpenSandbox. Four distinct approaches represent browser automation no longer having a single dominant solution — instead forming a differentiated infrastructure layer with clear division of labor. This typically signals a technology domain entering maturity.
1-Bit LLM Edge Inference: The Community's Most Grounded Skepticism
The discussion triggered by BitNet (370 HN points) reveals the real expectation gap in the developer community around "LLM democratization": there remains a significant distance between theoretical breakthroughs in model compression and practical usability. The most central question in 169 comments — "at what tasks is a 1-bit model already good enough" — will directly determine where the feasibility boundary for edge AI applications lies.
