Snyk Moves Internal Support Agent to Customer-Facing Product with 85% Ticket Deflection

Snyk has transitioned its internal support tool, Snyk Assist, into a core customer-facing feature available to all paying users. Originally built to help the internal support team triage thousands of cases, the agent now sits in the top bar of the Snyk web UI, allowing customers to find answers, check vulnerabilities, and open support cases directly. The move follows a phased rollout that began with internal testing, moved to the support portal in April 2026, and reached the core product in September 2026.

Architecture and Security

The system runs on a single LangGraph runtime, managed via LangSmith for observability, and is accessed through multiple surfaces including a Slack app, web app, and API. To meet the strict security standards required for a security platform, the team implemented several key design decisions:

• Permissions-aware tools: Tools are registered per user at request time, ensuring the agent can only access data the signed-in user is already authorized to see.
• Middleware integration: Context management, guardrails, and model fallbacks are handled via LangChain middleware, allowing for easy updates without rewriting the agent loop.
• Persistent state: Conversation history is stored in PostgreSQL using a checkpointer, enabling multi-turn memory and seamless scaling across pods.

Evaluation and Impact

Snyk employs a rigorous evaluation strategy using LangSmith to ensure reliability. This includes offline evaluations with known good answers and automated red-teaming exercises, as well as CI gates that block pull requests if performance thresholds are not met. Online evaluations grade every production run to measure deflection rates and categorize user queries.

Since its launch to customers in April 2026, Snyk Assist has achieved significant operational metrics:

• 60,000+ queries handled.
• 500+ customer accounts engaged.
• 85%+ of sessions resolved without creating a support ticket, saving hundreds of hours for the support team.
• 250+ cases auto-detected and escalated to the appropriate teams.