Snyk Turns Internal Support Agent Into Customer Feature

Snyk has transitioned its conversational AI agent, Snyk Assist, from an internal support tool into a core customer-facing feature. Built using LangChain and LangGraph, the agent was first tested internally by Snyk's support staff to manage thousands of bi-weekly cases. In April 2026, Snyk deployed the assistant to its customer support portal, and on September 1, 2026, integrated it directly into the main Snyk product interface. The tool has now handled more than 60,000 customer queries, resolving over 85% of sessions without requiring a human support ticket and automatically escalating more than 250 cases to the correct teams.

To scale the agent safely, Snyk's engineering team designed a single LangGraph runtime that powers multiple surfaces, including a Slack application, a web app, and direct API access. The architecture relies on PostgreSQL to persist conversation history and manage state across pods. Security is maintained by dynamically attaching tools at request time based on the active user's specific permissions, ensuring the agent never exposes restricted data. Furthermore, Snyk utilizes LangChain middleware to handle context, guardrails, and model fallbacks, allowing developers to modify agent behavior with simple configuration changes rather than code rewrites.

For AI practitioners, Snyk's rollout demonstrates a robust framework for moving LLM applications from prototype to production. The team managed risk by using LangSmith to trace every model and tool call from day one. They established a continuous integration gate where every pull request runs the agent against offline evaluation suites, including automated red-teaming exercises and accuracy checks. By grading production runs with scheduled online evaluation jobs, Snyk can map customer difficulties in real time and quickly convert problematic traces into new training datasets using the LangSmith Model Context Protocol server directly inside their IDEs.