Towards a Disaggregated Agent Filesystem on Object Storage (penberg.org)

A new filesystem abstraction called AgentFS has been proposed to enhance AI agents' interaction with the file system, moving beyond traditional models that rely on local disk files. Built on a SQLite database, AgentFS enables a portable and queryable environment where agents can access filesystem operations while maintaining state across ephemeral compute, essential for increasingly complex multi-agent workloads. This innovation addresses the need for seamless snapshots, state portability, and observability during agent operations, creating a unified single-file context for debugging and collaboration. The significance of AgentFS lies in its promise to disaggregate the agent filesystem through a combination of SQLite's write-ahead log (WAL) architecture and object storage solutions like Turso. By allowing agents to function in serverless environments and enabling state synchronization across machines, this hybrid approach facilitates real-time collaboration and minimizes infrastructure management challenges. Key technical implications include automatic conflict resolution during multi-agent interactions, enhanced historical state tracking through WAL, and future potential for improving write concurrency and checkpointing efficiency. This advancement represents a critical evolution for the AI/ML landscape as agents scale and require more dynamic storage solutions.