This is not a list of demo projects. This is not a list of repositories that got posted to Hacker News once and died. Every project here is actively used, patched, extended, and deployed by production ML engineering teams right now. None of them are asking for funding. All of them solve real, specific pain points that every team building AI systems hits this year.

The quiet shift in AI engineering

Everyone is still arguing about model benchmarks. No one is talking about the thing that actually determines if you can ship working AI: the infrastructure around the model. Over the last 6 months the entire field has pivoted. The bottleneck is no longer model capability. It is harnessing that capability reliably. The best teams are no longer fine tuning models. They are building harnesses, planning layers, tool interfaces, and verification loops. All of the work described below is building that layer.

cs249r_book: The AI engineering curriculum that everyone is stealing from

Harvard Edge's open source textbook is not just a course. It is the first serious attempt to establish AI engineering as a formal discipline, separate from both machine learning research and software engineering. The project hit 42k stars in 3 months. Almost every production ML team has at least one engineer working through this curriculum. More importantly, almost every new agent framework and inference engine copies the mental models laid out here. It does not teach you how to train a transformer. It teaches you why your 4000 GPU training run stalls. It teaches you how KV cache actually maps to silicon. It teaches you the physics that no tool or cloud provider will ever tell you about. The curriculum includes TinyTorch, a 20 module exercise where you build an entire ML framework from scratch. You do not understand a system until you have built one. This is the only educational resource that actually enforces that rule.

awesome-harness-engineering: The discipline no one told you about

Harness engineering is the practice of building the scaffolding around an AI agent. It is the single most important skill for production AI right now, and almost no one is teaching it. This curated list is the canonical reference for the field. It collects every public paper, pattern, failure mode and reference implementation for building reliable agent harnesses. The core insight here is brutal and correct: 90% of agent failures are not model failures. They are harness failures. You can take exactly the same model and move it 20 positions up or down any benchmark just by changing the harness around it. If you are building agents and you have not read this list, you are reinventing 30 different mistakes that other people have already documented and fixed.

planning-with-files: The agent skill that shipped first

This is the single most installed agent skill in the world right now. It runs on Claude Code, Codex, Cursor, Copilot, Gemini CLI and every other agent runtime. It has 19k stars and was forked 1100 times in 3 weeks. It does one extremely simple thing: it writes task_plan.md, findings.md and progress.md to disk. That's it. That single change fixes 70% of all common agent failure modes. Agents no longer forget what they were doing. They survive context clears, crashes, and restarts. They can be paused and resumed days later. They do not wander off task. Meta acquired Manus for $2 billion for exactly this capability. This repository implements it correctly, in open source, for every agent runtime.

Every agent team in the world is running some variant of this pattern right now. Most will not admit it.

awesome-claude-code: The unofficial agent ecosystem

Anthropic released Claude Code 3 months ago. They have not published any official documentation for extension development. This community curated list has become the de facto standard reference. It documents every hook, event, parameter, anti-pattern and working plugin that exists for Claude Code. At this point it is far more complete and accurate than Anthropic's official docs. The quality of entries here is unusually high. Every submission requires working code, reproduction steps, and a threat model for anything that touches security.

Category	Number of maintained entries
Core Documentation	12
Skills	47
MCP Servers	21
Hooks	18
Runtimes & Adapters	9
Security & Sandboxing	7

antigravity-sdk: Google's quiet agent runtime

Google released this SDK with no announcement. No blog post, no press release, just a repository pushed to GitHub. It is the most well designed agent runtime available today. It properly abstracts the agent loop, state management, tool registration, policies and streaming. It supports MCP natively. It has proper async interfaces. It does not leak implementation details. Most importantly: it works. There are no broken demos, no unimplemented methods, no marketing fluff. This is production code that Google runs internally, released for anyone to use. If you are building an agent application today, this is the runtime you should be evaluating.

claude-video: Give agents eyes

This skill lets Claude watch any video. Local files, YouTube, Loom, TikTok, anything. It extracts scene aware frames, pulls transcripts, deduplicates static frames, and hands everything to the model correctly. It solves a problem that every single person using agents has hit: you paste a link to a screen recording and the agent just guesses from the title. This is the first general purpose perception layer that actually works reliably. It installs in one command on every agent runtime. It has already been adopted by every major support and QA team that uses AI agents.

nano-vllm: Inference you can actually read

vLLM is the standard inference engine for almost everyone. It is also 100,000 lines of extremely optimized, extremely unreadable code. No one outside the core team understands how it actually works. nano-vllm implements every core optimization from vLLM: prefix caching, paged attention, tensor parallelism, CUDA graphs. It does it all in 1200 lines of clean, commented Python. It is also faster. On an RTX 4070 it beats upstream vLLM by 5% on identical workloads. This is not a toy. This is a reference implementation that you can modify, debug, learn from, and ship. For small deployments many teams are already replacing upstream vLLM with this.

Inference Engine	Throughput (tokens/s)	Lines of code
vLLM 0.7.1	1361.8	~112,000
nano-vllm 0.1	1434.1	1217

TradingAgents: Multi-agent done right

Almost every multi-agent demo is fake. They work for the video and then fall apart on any real input. TradingAgents is the exception. It implements a full structured multi-agent workflow modelled after actual trading firms. Separate agents for fundamentals, sentiment, technical analysis, research, risk and portfolio management. Agents debate, disagree, and produce verifiable outputs. It supports every LLM provider. It has proper checkpoint resume. It persists decision history and learns from past mistakes. This is not a trading bot you should run with real money. It is the best public reference implementation for structured multi-agent workflows that exists. If you are building any kind of multi-agent system you should copy this architecture.

ComfyUI: The quiet production workhorse

ComfyUI is the most underrated project in all of AI. Everyone thinks it is just a UI for generating images. It is actually the most mature general purpose workflow engine for AI that exists. It runs on every GPU, every accelerator, every operating system. It has proper incremental execution, checkpointing, versioning, and a massive ecosystem of extensions. It has a stable API that never breaks. More and more teams are abandoning agent frameworks and just building their workflows as ComfyUI graphs. It is not fashionable. It just works.

What comes next

All of these projects point in the same direction. The era of building monolithic models is over. The era of engineering systems around models has just started. None of this work will get covered in press releases. None of it will get 100k retweets. But this is the code that will actually power every production AI system built over the next 3 years. If you want to work on the important stuff, stop arguing about benchmark scores. Go contribute to one of these repositories.