Files over tools: how we built the Knock Agent using a virtual file system and bash

In March 2026, Knock introduced the Knock Agent, an AI tool designed to manage customer messaging resources by creating workflows, templates, and user segments, and assessing messaging performance. This agent operates through the Knock dashboard, Slack, the API, or MCP server, and was developed using bash, a virtual file system, and Knock's management API. The development team aimed for a scalable solution, ultimately building a virtual bash environment in Elixir, which allowed the agent to utilize a filesystem instead of multiple tools to access and modify data. For observability, they implemented OpenTelemetry traces and a homegrown evaluation system. The system supports learning new skills, including customer-defined ones, by accessing information via a CLI and a debugging skill, allowing it to adapt to various resource types beyond its initial focus on workflows and templates. The agent is intended to be efficient, avoiding performance costs associated with a full sandbox and instead utilizing a virtual, in-memory option. Future plans include refining session summarization, expanding evaluation cases, and exploring subagents for specialized tasks, ensuring the agent remains adaptable and responsive to evolving customer needs.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.