Mesa: A Versioned Filesystem for Agents
Mesa is a POSIX-compatible, versioned filesystem designed to support enterprise AI agents that create, edit, and manage long-lived documents requiring audit trails, human approvals, and parallel execution. It combines the durability and access controls of code repositories like Git with the simplicity of a filesystem interface, enabling agents to read and write files without managing versioning manually. Unlike cloud storage or traditional Git systems, Mesa handles agent-scale workloads, supports fine-grained permissions, and allows sparse file loading for efficiency. It can be mounted via FUSE or integrated through SDKs, making it adaptable across sandboxes and runtime environments.
Opening excerpt (first ~120 words) tap to expand
Imagine you’re building a legal-tech agent that can help with real-estate transactions. The v1 was a simple chat-with-the-docs app implemented with a RAG pipeline and some LLM calls chained together using some framework like LangChain or AI SDK. Now you want to build a v2 that can execute on more critical tasks such as drafting LOIs, redlining contracts, and co-ordinating signatures on documents. You reach for a pre-built agent harness like the Agent SDK, spin up a sandbox, and make fast progress until you hit a wall: How do you persist your documents beyond the lifetime of your ephemeral sandbox? How do you give your agent fine-grained control over which documents it can use for context and which documents it's allowed to edit? How do you build approval workflows for sensitive edits and…
Excerpt limited to ~120 words for fair-use compliance. The full article is at Mesa.
Discussion
0 commentsMore from Mesa
