Sam: So Google's now handing out pre-packaged AI agent skills... like file editing and web search.
Kai: And it's based on an open standard! That's the exciting part.
Sam: Right. I'm Sam.
Kai: I'm Kai. And this is Trending Repos for Tuesday, November 17th, 2026.
Sam: We're tracking three big stories from open source today, and they're all about the wild west of AI agents.
Kai: We'll start with that Google project, `stitch-skills` — it's a new plug-and-play library for giving your AI agents superpowers.
Sam: And Tencent is jumping in with `LoopForge`, a self-hosted workflow engine that—and this is key—puts a human in the loop.
Kai: And after that, we'll look at how you know if your shiny new AI skills are any good. NVIDIA thinks it has the answer: `SkillEvaluator`.
Kai: Okay, let's start with Google. The `google-labs-code` org just dropped `stitch-skills`.
Sam: Okay, so break it down. What exactly is an 'Agent Skill' in Google's terms?
Kai: Think of them as pre-built capabilities you can give to an AI. The repo already includes skills for file editing, web search, and—get this—shell command execution.
Sam: Wait, hold on. You're giving an AI direct, programmatic access to your shell? What could possibly go wrong?
Kai: It's about having a standard, Sam! That's the whole point. These skills follow the Agent Skills open standard, so any compatible agent—think Antigravity or the Gemini CLI—can use them interchangeably.
Sam: Standardized vulnerabilities, you mean. If there's a flaw in the shell skill, now every agent that uses it is a potential backdoor.
Kai: Or standardized power. So, here's what you could build with this. Say you're making a custom coding assistant. Instead of teaching it how to mess with the file system, you just import the `file_edit` skill. Boom. Your agent now knows how to read, write, and modify files.
Sam: And the first step is...?
Kai: Just clone the `stitch-skills` repository from GitHub. The whole library of skills is right there.
Sam: So Google is handing out LEGO bricks for building autonomous agents. I'm just saying—check those bricks carefully before you give them the keys to your filesystem.
Sam: Alright, let's move on to orchestration. Tencent just released `LoopForge`.
Kai: Orchestration... yeah, that feels like the next battleground. So this isn't about one agent, but managing a whole team of them?
Sam: Exactly. It’s a self-hosted system for creating workflows that involve multiple coding agents—it specifically mentions CodeBuddy, Codex, and Claude Code. But the most important feature is 'human approval gates'.
Kai: Ah, the sanity check! So the AI proposes a change, and a human has to sign off before it gets executed? After that last story, I am all for this idea.
Sam: It's essential, really. You chain tasks across different agents, but with a pause for a human to say, 'Yes, proceed,' or 'Whoa, terrible idea, stop.'
Kai: Okay, I really see the value there. So, what could you build with this?
Sam: Imagine a code-generation pipeline. Agent One drafts a new function based on a ticket. LoopForge then pauses the workflow and assigns a pull request to a senior developer. Once they approve, it triggers Agent Two to write the unit tests for that new function.
Kai: That's a killer application. You get the velocity of AI development without the terror of an agent going rogue and deleting production.
Sam: And since it's self-hosted, your code and workflows stay on your own infrastructure. To start, you just clone the `LoopForge` repo and follow the instructions to get the server running.
Kai: A self-hosted, multi-agent orchestrator from Tencent... That's a serious move.
Sam: It's a land grab. Everyone wants to own the platform that manages the agents.
Kai: Alright, that brings us to NVIDIA's `SkillEvaluator`, which feels like the perfect capstone to this whole conversation.
Sam: It really is the missing piece. So you've used Google's library to build a skill, and Tencent's tool to orchestrate it. But... is the skill any good?
Kai: Exactly! `SkillEvaluator` is a testing framework for AI agent skills. NVIDIA calls it a 'multi-tier framework,' which seems to cover the whole lifecycle of a skill.
Sam: What do they mean by 'multi-tier'? Are we talking unit tests, integration tests...?
Kai: Basically, yes, and then some. It has quality gates, it can detect if two skills are redundant, and—this is huge—it can generate synthetic evaluation datasets.
Sam: Whoa, synthetic data gen is a big deal. So you don't need to manually create thousands of test cases to get good coverage?
Kai: Right! And it even supports 'live agent evaluation' to measure how skills perform as part of a running agent, not just in isolation.
Sam: So it's like a full-blown QA department for your AI's abilities. I love it.
Kai: Totally. So for the 'what you could build' part: Say you've made a skill to refactor legacy Java. You point `SkillEvaluator` at it, and it generates thousands of synthetic Java files to test its accuracy, performance, and safety... before it ever touches your production code.
Sam: That's the kind of safety rail I've been waiting for. How do you get started?
Kai: It's the same drill: clone the `SkillEvaluator` repository from NVIDIA and follow the docs to set up the environment.
Sam: So the theme today is clear: Build your agent's skills, orchestrate them with human oversight, and for everyone's sake, test them rigorously.
Kai: Alright, let's do a quick recap.
Sam: Google's `stitch-skills` gives you open-standard building blocks for your agents.
Kai: Tencent's `LoopForge` helps you orchestrate them, self-hosted, with a human in the loop.
Sam: And NVIDIA's `SkillEvaluator` is there to test if those skills are actually safe and effective.
Kai: Before we go, quick one: did you see that Terminal UI apps are having a moment again? The `charmbracelet` ecosystem is on fire.
Sam: Kai, everything old is new again. TUIs are great. They're fast, they're lightweight... just don't try to sell me on a new text editor.
Kai: Deal. It's just cool to see so much innovation happening right there in the command line.
Kai: And that's our show! Thanks for listening to Trending Repos.
Sam: We'll be back tomorrow. Until then, maybe don't give your new AI agent shell access on the first date.
This show is made with AI: the hosts’ voices are synthetic and the scripts are AI-assisted. Every story links to its original source.