According to God of Prompt on Twitter, CSIRO Data61 and ArcBlock published a software architecture paper proposing that AI agents treat memory, tools, knowledge, and human input as a mounted filesystem that agents browse at runtime instead of preloading a large context window at boot. According to the tweet source, the approach reframes agent I O as filesystem operations, enabling on-demand retrieval that can reduce token costs and latency in production agents. As reported by the originating tweet, the paper is positioned as systems architecture rather than ML research, suggesting near-term adoptability for enterprise agent platforms, RAG pipelines, and tool-augmented workflows. According to the tweet, this design could standardize interfaces for external tools and knowledge bases, improving observability, access control, and compliance by leveraging familiar filesystem semantics. According to the tweet, the proposal addresses current bottlenecks in long-context models by shifting from static prompts to runtime browsing, a change that could enhance reliability, debuggability, and modular scaling in multi-agent systems.