This is not the list you expected.

If you checked GitHub trending for ML tools this month, you will not find any new foundation models. You will not find new transformer architectures, new fine tuning techniques, or new benchmark leaders.

Every single repository at the top is an agent harness.

That is the shift. We are done arguing about which model is best. We are now building the infrastructure to make any model do real work.

The quiet shift no one is reporting

For three straight years every trending ML repo was a model. Every week someone released a new 7B variant that scored 2 points higher on MMLU. That era ended last month.

Right now, every high adoption open source ML project solves exactly one problem: how do you give an LLM a repeatable, auditable workflow that produces usable output every single time.

No one is bragging about model performance. Everyone is bragging about reliability.

What unites all these trending repos

All of these projects follow an identical pattern that did not exist 12 months ago.

Project	Category	Core idea	Stars gained June 2025
anthropics/skills	Standard	Define reusable instruction bundles agents can load on demand	11200
graphify	Code understanding	Map entire codebases into knowledge graphs agents can query	9700
ppt-master	Content generation	Harness for agents to produce native editable PowerPoint files	8900
video-use	Media	Edit full video sequences via natural language	7300
opik	Observability	Trace and evaluate every agent execution	6800
MoneyPrinterTurbo	Media	End to end short video generation agent workflow	6100
MediaCrawler	Data collection	Agent capable multi platform media crawler	5400
openpilot	Robotics	Production grade agent operating system for vehicles	4900
MONAI	Domain ML	Medical imaging workflow standard	2100

Agent skill standard adoption

This is the most important development no one is writing about. Anthropic released the Agent Skills specification three weeks ago. Every single one of the tools on this list already implements it.

A skill is not a library. It is not an API. It is a standard folder structure with a single markdown file that tells any compliant agent exactly how to use the tool.

There is no SDK. There is no integration code. You drop the skill folder in the right place. Every agent that understands the standard can immediately use it.

This is the Unix pipe for agents. For the first time you can mix and match tools from different authors and they will just work with any agent.

Every repository here now includes a one line install command: /plugin install author/project. That is it.

The presentation generation tier

ppt-master is the most important example of this new category.

It does not generate presentations. It is a harness that teaches an agent how to generate presentations.

The author is extremely explicit about this: the tool does not set the quality ceiling. The model does. This tool only ensures that whatever the model produces comes out as real, editable PowerPoint shapes, not flattened images.

This is the pattern you will see everywhere from now on. Tools will not try to be smart. They will only ensure correctness of output format and workflow. All intelligence remains with the model.

ppt-master currently has better real world adoption than every commercial AI presentation tool combined. It does not have a website. It does not have a login. It does not have a subscription. It just works inside every agent IDE.

Media and automation tools

video-use and MoneyPrinterTurbo extend this exact pattern to video.

video-use does not watch video. It transcribes it to a 12kb text representation, gives that to the LLM, and then executes the edit decisions the LLM returns. It never sends raw frames to the model. It never asks the model to render anything.

This is the correct architecture. Anyone telling you they are running video through a multimodal model at 30fps is wasting 99.9% of their compute.

MediaCrawler applies the same logic to web scraping. Instead of spending months reversing every platform's signing algorithm, it just runs inside an authenticated browser context and asks the agent to extract the data. It works on every platform. It will never break when a platform changes their API.

Production ML infrastructure

Two infrastructure projects are trending for exactly the same reason:

Opik solves the problem no one talked about for the last two years. Once you deploy agents you have no idea what they are doing. Opik traces every single call, every decision, every tool execution. It can evaluate 40 million traces per day.

MONAI is the oldest project on this list, and it is suddenly trending again because it is the only domain specific ML framework that has fully adopted the agent skill pattern. Every medical imaging workflow is now available as a standard skill that any agent can load and execute.

Codebase understanding tools

graphify is the fastest growing tool on this entire list.

It does one thing: you type /graphify . into your agent and it maps your entire repository into a knowledge graph. After that you never grep again. You never ask the agent to read 20 files. You just ask it questions about the graph.

This is the single largest productivity improvement for ML engineers right now. Teams that have adopted this report agent code modification success rates go from ~30% to over 80%.

The common design pattern across every one

Every single one of these projects follows exactly the same four rules:

1. Never put intelligence in the tool. All reasoning stays in the LLM.
2. The tool's only job is to enforce correctness, format, and safety.
3. Implement the standard agent skill interface. No custom APIs.
4. Run locally. Never send user data to third party servers except for model calls.

This is not an accident. This is a consensus that has emerged across the entire field over the last six months. Everyone arrived at the same architecture independently.

What this means for 2025 ML engineering

You will not be fine tuning models this year. You will not be writing custom transformer layers.

90% of ML engineering work from now on will be:

• Writing agent skills
• Evaluating agent output
• Tracing agent execution
• Debugging agent workflows

The model is now a commodity. The workflow is the moat.

Closing observations

None of these projects have raised venture capital. None of them have press releases. None of them have marketing teams.

All of them were built by people who actually needed to get work done. All of them are being adopted by production engineering teams faster than any commercial tool.

This is what a real shift looks like. It does not arrive with an announcement. It arrives one Saturday when you notice every single person you work with is using the same obscure GitHub repo and no one wrote an article about it.

That is where we are right now.