Disclaimer: This text is also generated by Claude sonnet 4.5
I had an idea: What if AI agents could be packaged and distributed like npm packages? Portable, versionable, and platform-independent.
The problem I'm solving:
• OpenAI's GPTs are locked to their platform
• LangChain requires heavy coding for custom agents
• No standard format exists for AI agent distribution
• Creators can't easily share or monetize agents across platforms
So I created the .sagent package format - a standardized way to package AI agents.
What it is:
• A ZIP file containing YAML config, system prompts, examples, and knowledge bases
• Portable across any LLM platform that supports it
• Versionable with semantic versioning (1.0.0, 1.1.0, etc.)
• Includes validation and safety checks
• MIT licensed, fully open specification
I've also built a reference implementation - a marketplace prototype where you can upload .sagent packages and chat with agents in real-time.
Tech stack:
• Python 3.12 + Flask 3.0
• gevent for async I/O (handling concurrent chats)
• Anthropic Claude 3.5 Sonnet API
• SQLAlchemy
Full transparency: This prototype was built entirely with Claude Code (Anthropic's AI coding assistant). I came up with the idea, but Claude wrote all the implementation code. This is both a test of AI-assisted development and a real product I want to build.
The .sagent format specification: https://github.com/jinnks/special-agents/blob/main/AGENT_PACKAGE_SPEC.md
GitHub: https://github.com/jinnks/special-agents
I'd love feedback on:
1. Is this package format useful? What would you change?
2. What other metadata should .sagent files include?
3. Security concerns with user-uploaded agent packages?
4. Better ways to handle agent knowledge bases?
5. Thoughts on AI-assisted development? (This entire codebase is AI-generated)
This is very much a prototype - I know it needs work. Looking for contributors who find this problem interesting.
What would you want in a standardized AI agent format?
reply