Files over tools: how we built our agent with a virtual filesystem and bash (knock.app)

In March 2026, Knock launched the Knock Agent, an AI-powered tool designed to streamline customer messaging management. This agent can create and update workflows, templates, and segmented audiences through various interfaces, including the Knock dashboard and Slack. The development team aimed to enhance the agent's functionality by employing a unique architecture that leverages a virtual filesystem and bash scripting, allowing the agent to dynamically access and manipulate customer data without overloading context management. The significance of this innovation lies in its ability to provide a more scalable and efficient way to manage messaging resources. Rather than relying on a tool-per-resource approach, which would complicate the agent's functionality, the team opted for a filesystem-based architecture inspired by a similar approach from Vercel. This shift not only streamlines the communication between the agent and its data sources but also aligns with a local-first philosophy that enhances user experience. By converting complex interactions into basic filesystem operations, the Knock Agent is better equipped to handle real-time queries and persist changes efficiently, ultimately paving the way for future enhancements and the introduction of more advanced capabilities such as subagents and improved evaluation systems.