The Folder Is the Agent
I'm running 44 AI agents across multiple projects. Each one is just a model pointed at a folder.
TL;DR
- The author experimented with multiplying AI agents for efficiency but found that more agents didn't necessarily increase speed, leading to an overwhelming number of results to evaluate.
- The realization was that a project folder, containing files like CLAUDE.md, skill definitions, and accumulated context, acts as an effective AI agent by making a general model a specialist.
- Currently, the author manages 44 agents, each operating within a specialized folder, with a custom dispatch layer routing tasks between them.
- This 'folder as agent' pattern allows a single model (like Opus 4.6) to act as different specialists (e.g., Rails engineer, ops engineer) simply by changing the folder it interacts with.
- A dispatch layer was built to manage these numerous agents, using slash commands like /hey for status reports and /orchestrate to initiate tasks, replacing manual terminal management.
- Challenges at scale include encoding bugs, context drift, and agent stalls, highlighting the need for careful planning and trust-building before full automation.
- The 'folder is the agent' concept is scalable, with the potential for further advancement through services like Claude Managed Agents, which handle background processes for specialized agents.