The filesystem has been organizing code since before Unix. Great for humans. Terrible for AI coding agents, according to Sumato AI. The San Francisco startup just open-sourced Spath and Splan, two grammars that attempt to move coding agents away from file operations entirely. Spath is a semantic addressing format. Instead of "src/services/user.go line 47", an agent writes something like "github.com/example/app/database.User.GetID". Splan handles the verb side: what to replace, what to remove, what to add. Both work at the symbol level, not the file level.
The philosophy here is what Sumato calls "narrative hygiene." LLMs think in stories. Every context window is a narrative. When an agent has to grep through files, parse paths, and recover from partial edits, that narrative gets messy. Jason Mooberry, who authored the announcement, puts it simply: tokens down, quality up. The Splan format is also transactional. If one operation fails, nothing applies. No half-mutated codebases, no expensive recovery narratives.
Founders Einar Vollset (formerly an exec at HubSpot) and Nicky Winestead argue that filesystems are human artifacts we've forced onto AI. Smalltalk and Intentional Programming tried moving programming into semantic space decades ago. Humans didn't bite. AI works at whatever abstraction you give it. The specs are released under CC BY 4.0, and Sumato explicitly wants them in public training data. They're betting that a shared semantic addressing layer becomes more valuable as more tools adopt it. Both specs are at version 0.1.0. The Go dialect is already in use at Sumato.