Files are the interface humans and agents interact with
Filesystems are being rediscovered as a powerful, durable interface for both humans and AI agents, with many arguing that plain files in standard formats (markdown, CSV, JPEG, EXIF) form a more resilient “database” than proprietary SaaS silos or opaque app-specific schemas. Commenters highlight how agents can use traditional Unix tools and directory structures to organize, search, and maintain long-lived knowledge, while still allowing more specialized databases to be layered on top when needed. Others raise concerns about security, metadata portability, mobile OS restrictions, and the growing presence of low-quality AI-generated writing, suggesting that current file-centric agent workflows are a useful but likely transitional stage.
Legacy use of filesystems as “databases”
- Several recall historic patterns: using directory trees and file names as indexes or key–value stores when RAM was scarce (e.g., early consoles, low‑memory systems).
- Commenters say we’re “back to the old ways”: LLM agents using files and directories as primary data structures.
Files, standards, and SaaS lock‑in
- Strong support for “boring,” open formats (JPEG, EXIF, markdown, CSV) as long‑term source of truth.
- SaaS apps and proprietary formats are criticized as short‑lived and fragile; they accumulate technical debt and risk losing data when services die.
- Some users now store everything as plain files and let tools/agents index or layer on top.
Photo management and metadata
- Files + EXIF as canonical archive is praised; libraries can be re‑indexed by new tools.
- Extended attributes and XMP sidecars are seen as fragile: not well standardized, easily lost when copying across media, and annoying to manage as multiple files per photo.
- Frustration that modern photo apps store edits/tags in external databases, breaking portability between services.
Filesystem vs databases and alternative models
- Many describe a filesystem as a simple database: tree+metadata, with backups via file copies and optional content hashes.
- Others call hierarchical trees a “terrible abstraction” and prefer relational or UUID‑based models with queryable attributes, generating views/directories on demand.
- Discussion touches on NTFS, ReFS, BeFS/Haiku, and Plan 9/9P for richer indexing, attributes, and namespace‑based security.
Agents, tools, and security
- Enthusiasm for agents that operate on local files using bash/CLI tools; agents benefit from unified, user‑owned file hierarchies.
- Counterpoint: this “everything is a file” agent model may be insecure and will need stricter permissioning, akin to app sandboxes.
- Some propose embedding narrowly scoped agents inside specific applications (word processors, spreadsheets) with task‑focused capabilities and organizational permission hierarchies.
Meta: AI content and article scope
- Mixed reactions: some found the piece clarifying; others were disappointed it wasn’t about new filesystem designs but yet another AI/agents article.
- Heated subthread on whether the article was LLM‑written, with calls for explicit labeling of AI‑assisted writing and complaints about rising “AI slop” online.