19 GitHub Repositories That Take AI Agents from Chat to Shipping
A tweet compiled 19 repositories covering skills, team, plan, memory, and tools. Pair that with the system breakdown of Grok Bot, and you'll see: the model is just the brain; the real difference is in the operating system.
A tweet compiled 19 GitHub repositories, with the caption: Skills → team → plan → memory → tools → ship. The original tweet is from @akarixbt and is titled “This is f*cking gold...”.
This ordering itself says one thing: you need skills, team, plan, memory, and tools first, and only then can you deliver. A single model is not enough.
Skills
- Matt Pocock's skills: Skills for real engineers, from his .agents folder. He explicitly says “not vibe coding.”
- Addy Osmani's agent-skills: Production-grade engineering skills; 26 skills cover the full lifecycle from spec to ship, with built-in anti-rationalization tables.
- superpowers: A skill framework + development methodology, from brainstorming to TDD to subagent-driven development.
- Vercel agent-skills: Vercel's official agent skills.
- awesome-agent-skills: 1000+ curated skills.
- npx skills: An open-source tool that installs skills into 70+ agents.
Team
- Lauren Tan's noodle: Orchestrating agents with skills.
- Gas Town: A multi-agent workspace manager.
- Garry Tan's gstack: 23 tools, with roles including CEO, designer, engineering manager, and release.
Plan
- BMAD Method: An agile, AI-driven development method.
- Spec Kit: A spec-driven development toolkit, official from GitHub.
- Task Master: AI task management for agents.
- AGENTS.md: An open format for instructing agents.
Memory
- Beads: An upgrade to agent memory.
- mem0: The memory layer for AI agents.
- Context7: Provides agents with up-to-date documentation.
Tools
- Composio: 1000+ toolkits with authentication.
- GitHub MCP: GitHub's official MCP server.
- Firecrawl: Search and scrape web pages, for agents.
The list is just parts. The real question is: how do you assemble them?
rari🈀 breaks down the system design of Grok Bot in “Grok Bot: The AI Team That Never Sleeps,” answering this question. Original link
Grok Bot is an agent launched by xAI: it has a persistent computer, real tools, persistent context, routines, and can coordinate with other agents. Grok 4.6 is the reasoning model. But according to rari🈀, the model is only the brain, not the whole thing. An agent system consists of several layers: harness (environment), loop (improvement loop), graph (coordination), and approval.
A few key points:
The difference between chatting and delivering. A chatbot returns answers; a Bot returns completed work. The output is no longer a paragraph saying “you should do this,” but a fixed CRM, a prepared draft, a reproduced bug, an organized folder. rari🈀 suggests that your first task should not be “run my company,” but rather “open the latest support report, find the three biggest changes, and leave a summary with citations in the review folder.” Small enough to verify within a minute.
Classifying failure modes. If a Bot forgets preferences, it's a state problem. If it calls the wrong tool, it's a routing problem. If it sends unapproved content, it's a permissions problem. If it repeats failing actions, it's a loop problem. If multiple Bots ask each other who does what, it's a graph problem. rari🈀 says better prompts improve only one run; better architecture improves every run.
The approval line is split by reversibility. Research, summarization, drafting, organizing, preparing, staging, and simulation can be done autonomously. Sending, publishing, purchasing, transferring money, deleting, changing permissions, modifying production, and accepting legal terms must stop before approval. An ideal outcome is: all reversible steps are complete, and all irreversible steps are clearly staged.
The shared computer is the security boundary. Multiple Bots share one user-scoped persistent computer, including browser sessions, files, and app logins. This is the foundation of collaboration, and it also means the security boundary is at the account level. Role descriptions cannot substitute for security controls. If two Bots require different trust levels, separate their accounts and environments.
Automation rots. Websites change, credentials expire, preferences drift. rari🈀 suggests having each Bot submit a receipt every week, including run count, pass rate, and failure reasons. Then spot-check one artifact yourself. If a routine could disappear without you noticing, it should be deleted.
These points are excerpted from rari🈀's article; read the original for details.
The 19 repositories are the parts; the Grok Bot breakdown is the framework. Next time your agent only chats and doesn't do the work, don't blame the model first. Look at skills, team, plan, memory, and tools—which layer is missing?
发布时间: 2026-11-22 12:21