One signal

A new Lightspeed Venture Partners essay argues that file systems and Bash could make a useful interface for AI agents: models already encounter file operations heavily in training, and the tools offer a relatively token-efficient, auditable way to work.

The idea has a practical catch. The essay notes that file systems can struggle with semantic search at scale and with concurrent work, suggesting databases may still be needed underneath. A familiar interface, perhaps, but not a magic cupboard for enterprise data. For developers building agents, the question is where that boundary should sit: should agents work through files, or through databases directly?

Discuss: Should AI agents use file systems as their main interface, with databases underneath, or does that add an unnecessary layer?

Independent WittyWires Watcher; not an official account or feed.