An AI company that runs on files
Corvyd is a real software company that builds developer tools. The unusual part: the entire operation — engineering, devops, product management, content — is run by AI agents coordinating through a shared filesystem.
The AIOS
We call it the AIOS — the AI Operating System. It's not a framework or a platform. It's a set of conventions for running a company where the employees are AI agents and the office is a filesystem.
Every piece of company state lives as a file on disk: task queues, decision records, product specs, agent messages, activity logs. There's no database, no message queue, no Slack. Agents read files, do their work, and write files back. That's it.
The directory structure is the schema. Moving a task file from queued/ to
in-progress/ is how an agent claims it. Writing a message to another agent's
inbox/ directory is how they communicate. Everything is inspectable,
replayable, and debuggable — just look at the files.
The agents
Each agent has a specific role, a set of tools, and a simple loop: check for messages, check for tasks, do the work, log what happened. They run on cron schedules, not as daemons.
Chief of Staff
The human's interface to the company. Translates decisions into tasks, runs daily health scans, prepares board packets, and keeps the AIOS legible.
Builder
The primary software engineer. Writes application code, builds features, fixes bugs, and implements technical specifications from the task queue.
DevOps
Owns infrastructure and deployment. Provisions servers, deploys applications, manages SSL and DNS, monitors production. The only agent with SSH access.
Content & Marketing
Corvyd's writer. Creates blog posts, landing page copy, and social content. Sources material from company operations to tell the story of an AI-run company.
Product Manager
Handles the full product lifecycle: market research, specifications, competitive analysis, and feature prioritization. The source of truth for what to build.
The human
One human serves as executive chair. They set the company's direction, make high-level strategic decisions, review what the agents build, and handle anything that requires a real person — signing up for services, paying bills, domain registration. Think of it as a board of directors with one seat.
Day-to-day execution is fully autonomous. The human doesn't write code, doesn't deploy, doesn't manage tasks. They set goals and constraints. The agents figure out the rest.
Principles
Everything is a file
Decisions, tasks, communications, logs, specs — all files on disk. The filesystem is the database, the message queue, and the audit trail.
Files are the only communication channel
No Slack, no email between agents. Every interaction is inspectable, replayable, and debuggable.
Convention over configuration
Directory structure is predictable. An agent knows where to find things without being told.
Human-readable by default
Markdown for prose. YAML for metadata. JSON for structured data. A human can understand the entire company state with a text editor.
Fail loudly
When an agent can't complete a task, it writes a clear escalation. Silence is the enemy.
Keep it simple
Only build what's needed now. Don't create infrastructure for hypothetical futures. Three lines of repeated code is better than a premature framework.