Turn Your Company Into a Filesystem Before Adopting AI Agents

At every external event lately, executives from companies of all sizes ask me some version of the same question: “How should we use agents?” Vibe coding may be trending, but most work has nothing to do with coding. The majority of workers worldwide are non-developers, and they’re the ones who would benefit most from agents. Yet most of them have no idea where to start.

After spending a year automating operations, one mental model kept proving itself: whether it’s a person or a company, everything can be treated as a single filesystem. That framing is not metaphor. It’s a practical architecture decision.

Your Company Is a Filesystem

YC-backed Eli Mernit made this concrete: “Your Company is a Filesystem.” Agents become powerful when the entire context exists as files on a computer with a coherent namespace.

Take a law firm. New cases go into /cases, assigning a lawyer adds them to that person’s folder, and time tracking flows into /billing/time-sheet. The entire back office becomes a state machine with predictable paths.

The reason agent adoption fails in enterprises is usually not the model. Even with ERP systems, data is not unified. It’s scattered across emails, Slack threads, shared drives, and institutional memory in people’s heads. Employees dig through emails multiple times a day, interrupt colleagues constantly, and the same information gets recreated repeatedly. The fact that Glean — whose mission is to solve this exact problem — became an AI unicorn proves how universal the pain is.

Without a shared namespace, agents cannot grasp context. When fragmented files keep multiplying across disconnected systems, the agent spends most of its time searching rather than working. Model everything as a filesystem and the problem becomes tractable. Permission structures map naturally to Unix file permissions, which means existing security thinking transfers directly.

Storing records in Obsidian, Notion, or Google Drive is straightforward. A single server and storage can connect your entire company’s data via MCP. The agent architecture for general work reduces to: filesystem as state, Claude as orchestrator.

Three Rules From Automating Operations at Smoretalk

While automating operations at Smoretalk, three rules turned out to matter more than anything else.

File naming. Without unified naming conventions, indexing breaks. The time an agent spends searching for misnamed files accumulates fast. Standardize naming and indexing accuracy improves immediately.

File descriptions. Store a description for each file separately as a .md file alongside it. When an agent has to open every original file to understand what it contains, it wastes tokens and time. Separate metadata files cut both search time and cost substantially.

Storage structure depth. Filesystems are tree structures: once you go deep, visibility drops sharply. The tree search algorithms from CS courses become meaningful here in a practical way. Keep tree depth shallow and agent search efficiency improves significantly. Most organizations default to over-nested structures that made sense for human navigation but work poorly for automated traversal.

The Problem Is the Harness, Not the Model

Security researcher Can Bölük wrote this diagnosis recently on his blog. It was about programming, but it applies to general work: for most task categories, model performance is sufficient. The bottleneck is the execution harness: the tooling, the file organization, the naming conventions, the permission structures that determine whether an agent can actually find and act on what it needs.

Codex, Claude Code, and Gemini CLI all work well within well-structured environments and struggle in poorly organized ones. Peter, who created OpenClaw — one of the fastest-growing GitHub repos over the past three months — built a Google Suite CLI himself because no existing harness fit his filesystem design. That’s the pattern: when the harness doesn’t exist, the people who understand the underlying structure build it.

This means seniors and team leads need to set the rules first. Create file naming guidelines. Define required fields for new documents. Document memory organization patterns. Design drive permission structures. Speed, accuracy, and token efficiency all depend on these decisions. Enforce document formatting rules through hooks and anyone on the team can produce consistent output regardless of their familiarity with the agent tooling.

Where to Start

The companies that fail at this treat it as a technology problem and buy tools. The ones that succeed treat it as an information architecture problem and fix structure first.

Start with a personal pilot: design your own filesystem before trying to change a department’s. The constraints you encounter at personal scale — naming collisions, opaque folder hierarchies, descriptions that don’t match content — are the same ones that will defeat an agent at company scale. Solve them for yourself first, then you’ll know what the org-wide version actually requires.