Hi HN, Matias Molinas here.
For the past few months, my collaborator Ismael Faro and I have been wrestling with a core problem: building autonomous AI agents is still too much like traditional software engineering. It's confined to those who can write complex code, manage infrastructure, and navigate intricate libraries.
This led us down a rabbit hole, asking a series of "what if" questions:
What if we stopped treating LLMs as just another API call and started treating them as the runtime environment itself?
What if markdown, with its simple structure, could be a legitimate programming language for defining agentic behavior?
What if we could create a system so accessible that domain experts in law, marketing, or science could build the exact automation they need without writing a single line of code?
Our exploration into these questions has led to LLMUNIX, an experimental, document-centric operating system. You can see the main project here:
https://github.com/EvolvingAgentsLabs/llmunix
In LLMUNIX, there is no underlying Python or TypeScript application to run. The "OS" boots when a powerful LLM is given a master prompt (LLMUNIX.md) and a file system of markdown documents.
The agents and tools are not classes; they are .md files that describe their purpose and logic in structured English. The LLM acts as the CPU, interpreting these documents to plan and execute tasks.
The most significant capability we've discovered is autonomous tool creation. When a user gives LLMUNIX a goal it has never seen before—like "analyze this legal contract for risks"—it doesn't fail if a "LegalAnalysisTool" doesn't exist. It recognizes the capability gap, writes the markdown definition for the new tool itself, registers it in its SmartLibrary.md (the "package manager"), and then immediately uses it to solve the problem.
The Potential Impact & Why We're Sharing This Now:
We believe this approach points to a future where we can democratize the creation of agentic systems. It shifts the paradigm from coding complex logic to simply describing it. This could empower a whole new class of creators. Furthermore, every execution trace from LLMUNIX is a perfect, structured dataset for fine-tuning LLMs to have these orchestration skills natively, creating a powerful feedback loop for building smarter models.
We've launched Evolving Agents Labs because we know these ideas are in their infancy. This is an alpha space for experimentation. We are not sharing a finished product, but an invitation to think out of the box with us. We want to explore new paradigms, and we believe the HN community is the perfect place to do that.
We invite you to test these ideas. The repositories are structured to make it easy:
llmunix: The flagship OS. The best place to start and see it all work.
framework-core: The minimal, foundational blueprint if you want to build your own OS from our principles.
agent-examples: High-impact, non-technical scenarios that showcase the autonomous "zero-to-tool" capability.
We're incredibly excited and a little nervous to share this. We'd love to hear your critiques, ideas, and what you think the limits of this "markdown as an OS" concept are.
Thanks for reading.