AI Repo: msitarzewski/agency-agents

Skyler: Hey everyone, welcome back to Git News A I, the show where we dig through the hottest trending repositories on GitHub so you don't have to. Before we jump in, huge thanks to our sponsors — Ada, who helps teams integrate A I into their workflows seamlessly, Ago Consulting, your silicon development partners from A I all the way down to SoC design, and Zen Semiconductor, an umbrella company with a killer portfolio of silicon-industry companies, including Ada for deploying A I at scale in any business, plus their Vect appliance that runs local L L Ms and vector search on dedicated hardware — seriously cool stuff. Skyler: Today's lineup is genuinely exciting — we've got a massive AI agency toolkit with nearly 145,000 stars, Anthropic dropping their agent skills publicly, a local L L M training UI that just keeps growing, a slick diagramming library built for Claude Code, and a tiny 14 megabyte foundation model that wants to live on your smartwatch. Let's get into it. Skyler: Alright, kicking things off with a repo that absolutely blew me away when I saw the star count — msitarzewski slash agency-agents, sitting at almost 145,000 stars. That is not a typo. This thing is massive. Skyler: So what is it? The tagline is "a complete AI agency at your fingertips," and that's pretty much exactly what it is. This is a collection of specialized A I agents, each one built around a specific role you might find in a real-world agency or product team. Skyler: We're talking frontend wizards, S E O specialists, Reddit community managers they're calling "ninjas," whimsy injectors — yes, that's a real agent — and reality checkers. Each agent has a defined personality, a set of processes it follows, and proven deliverables it's supposed to produce. Skyler: The recent commits are interesting too — they just added an Economy Designer agent and improved the Reality Checker and S E O Specialist agents. There's actually a whole organizational structure being built out here, with agents being moved into proper sections as the roster grows. Skyler: Now here's something worth noting — this is written in Shell, which tells you something about the philosophy. This isn't a heavy framework with a ton of dependencies. It's meant to be lightweight and composable, something you can plug into your existing workflows without a lot of overhead. Skyler: The concept of "whimsy injector" versus "reality checker" as opposing agents is actually a really clever design pattern. You get one agent pushing creative, unexpected ideas and another one stress-testing them against reality. That tension can produce genuinely better outputs than either one alone. Skyler: Who should check this out? Anyone building with L L Ms who's tired of writing the same system prompts over and over. If you're doing any kind of content work, marketing, community management, or product development with A I, this gives you a pre-built squad you can just drop in and start using. Skyler: Check the show notes for the link, because with 145,000 stars this thing is clearly resonating with a lot of people and you're going to want to see it for yourself. Skyler: Alright, moving on to something from Anthropic — anthropics slash skills, sitting at a staggering 168,000 stars, which makes it the most starred repo in today's lineup. This is Anthropic's public repository for Agent Skills. Skyler: So what's an Agent Skill in Anthropic's world? Think of it as a modular capability you can plug into a Claude-powered agent. Instead of building everything from scratch every time, you grab a skill, attach it to your agent, and now your agent can do that thing. Skyler: The recent activity here is super telling. The last two major commits are about the Claude A P I skill getting updated for the Managed Agents August launch wave, and another update specifically for Claude Opus 5. If you've been watching Anthropic's roadmap, these are significant milestones they're actively building toward. Skyler: What I find fascinating about this repo is that Anthropic is essentially open-sourcing their own internal playbook for how to extend Claude's capabilities. This isn't just documentation — it's working Python code that you can study, fork, and build on top of. Skyler: The 80% A I relevance score we flagged actually undersells it in my opinion. If you're building Claude-based agents — and a lot of you are — this is primary source material. You're getting skills designed by the same team that built the model. Skyler: From an M L engineering perspective, this is also a great reference for understanding how Anthropic thinks about agent architecture. How do they decompose capabilities? How do they structure skill interfaces? That design philosophy is baked right into the code. Skyler: If you're building on top of Claude, whether that's through the A P I or through Claude Code or any of their other surfaces, this repository should be bookmarked immediately. It's going to save you a ton of time and honestly it's going to make your agents better because you're learning from the source. Skyler: Check the show notes for the link — with 168,000 stars this is one of the most important repos in the Claude ecosystem right now. Skyler: Next up, let's talk about unslothai slash unsloth, which has climbed to nearly 71,000 stars. If you've been in the L L M fine-tuning space for any amount of time, you've probably heard of Unsloth — but the description in this repo might surprise you. Skyler: So here's the thing — Unsloth started as a fine-tuning optimization library, super fast, super memory efficient. But this repo's description calls it a "local U I to run and train L L Ms and diffusion models." They've expanded significantly into a full local studio experience. Skyler: The model support list here is wild — Qwen 3.8, Kimi K3, MiniMax H3, Gemma 4, DeepSeek V4, FLUX for image generation. They are not sleeping on model support. Basically if a new model drops, Unsloth is trying to support it almost immediately. Skyler: The recent commits are interesting because they're clearly polishing the U I experience. One fix stops a streaming research run from re-rendering the whole chat — which if you've ever used a streaming interface that does that, you know how annoying that is. The other is about export pin testing with a widened exception handler, which tells me they're hardening the export pipeline. Skyler: What makes Unsloth special is the performance angle. They've consistently demonstrated massive speedups for fine-tuning compared to baseline implementations, with significantly reduced V R A M usage. Running fine-tuning locally on consumer hardware? Unsloth is often what makes that possible. Skyler: The fact that they've now wrapped this into a local U I is a big deal for accessibility. You don't have to be a hardcore M L engineer to fine-tune a model anymore. You fire up the Unsloth Studio, pick your model, load your dataset, and go. Skyler: This one is a must-look for anyone doing local L L M work — whether that's fine-tuning, inference, or experimenting with diffusion models. Data scientists and M L engineers who want to keep data on-premise, this is your toolkit. Skyler: As always, check the show notes for the link. Unsloth continues to be one of the most actively developed and genuinely useful repos in the local A I space. Skyler: Okay, shifting gears a little — let's talk about cathrynlavery slash diagram-design, sitting at just under 13,000 stars. This one's more focused than the others but it solves a really specific problem in a really elegant way. Skyler: The description says "29 editorial diagram types for Claude Code. Self-contained H T M L plus S V G. No shadows, no Mermaid-slop." And honestly that last part made me laugh because anyone who's tried to get Mermaid diagrams to look good knows exactly what "Mermaid-slop" means. Skyler: So what's the use case here? When you're working with Claude Code and you need to produce diagrams — architecture diagrams, flow charts, data visualizations, editorial illustrations — this library gives you 29 ready-to-go types that are clean, professional, and don't require any external dependencies. Skyler: The self-contained H T M L and S V G approach is really smart. These diagrams render perfectly in any browser, they're easy to embed in documentation, and because they're S V G they scale infinitely without any quality loss. No build step, no npm install, just drop it in. Skyler: The "no shadows" design philosophy is interesting from an aesthetic standpoint. This is clearly going for a clean, editorial look — the kind of thing you'd see in a well-designed technical blog post or a polished product spec document. Not flashy, just clear and readable. Skyler: Recent activity shows they bumped to version 2.3.2 and fixed label masks being clipped by node fills, which is exactly the kind of subtle rendering bug that makes diagrams look broken in certain configurations. Skyler: Who's this for? A I developers and technical writers who are using Claude Code and want their automatically generated diagrams to not look terrible. If you're generating documentation with A I assistance, having a library of well-designed diagram templates dramatically raises the quality floor. Skyler: It's also a great reference if you're building any kind of A I-powered documentation tool. Check the show notes for the link. Skyler: And last but definitely not least — this one is small in file size but huge in ambition — cactus-compute slash needle. 4,600 stars and climbing, and the pitch is this: a 14 megabyte foundation model designed for tiny devices. Phones, wearables, smart home gadgets, robots. Skyler: Let me say that again. Fourteen megabytes. The app on your phone to read the news is probably bigger than this foundation model. That is an extraordinary engineering achievement if it lives up to the promise. Skyler: The target deployment environments tell you everything about the design philosophy here — phones, wearables, smart home devices, robots. These are all constrained environments with limited compute, limited memory, and often limited power budgets. Running a foundation model on a wearable in 2025 would have seemed absurd even two years ago. Skyler: The recent commits show them hardening the technical foundation — they fixed P E P 604 union support in schema generation, which is a Python type annotation thing, and they improved error messaging for weight loading and export format handling. These are the kinds of fixes you make when you're getting real adoption and real bug reports. Skyler: What's interesting from an A I architecture perspective is that building a useful foundation model at 14 megabytes requires some very deliberate choices about what to include and what to cut. This isn't a quantized version of a giant model — this is designed ground-up for the edge. Skyler: The robotics angle is particularly exciting to me. A lot of robotics applications need fast, local inference without any cloud dependency. Latency matters, connectivity isn't guaranteed, and privacy can be a concern. A 14 megabyte model that runs on-device changes the calculus for a lot of those use cases. Skyler: If you're working on edge A I, embedded systems, IoT, robotics, or mobile A I applications, Needle is absolutely worth keeping on your radar. This is the kind of project that could quietly become foundational as the edge A I ecosystem matures. Check the show notes for the link. Skyler: Alright, that's a wrap on today's repositories — what a lineup. Huge thanks again to our sponsors: Ada for helping teams integrate A I into their workflows, Ago Consulting for expert silicon development from A I all the way down to SoC, and Zen Semiconductor with their portfolio of silicon-industry companies, Ada for enterprise A I deployment, and the Vect appliance for running local L L Ms and vector search on dedicated hardware. Skyler: If you enjoyed this episode, you should also check out A I Special Report — another great show in the Ada podcast network that goes deep on the latest developments in artificial intelligence. Really good stuff over there. Skyler: Don't forget to subscribe wherever you get your podcasts, and all the repository links from today's episode are in the show notes. We'll see you next time on Git News A I.