Snyk says its AI support agent resolves over 85% of sessions without a ticket

Snyk’s AI support agent is now inside its core product, rather than confined to a separate support portal. In an October 8 case study published by LangChain, Snyk details the September 1 rollout to every paying customer. It also reports more than 60,000 customer queries handled since April, with over 85% of sessions resolved without a support ticket.

A year inside support before customers got access

The starting problem was scattered information. Customers’ answers lived across documentation, support articles, release notes, learning materials and account systems. Meanwhile, Snyk’s support team handled thousands of cases every couple of weeks, each requiring classification and routing by product, severity and owner.

Snyk Assist spent roughly a year serving only the company’s support staff. Snyk says that kept early errors internal and exposed edge cases missed by offline tests. Customer access followed in the support portal in April 2026. An August 18 early-access announcement preceded the September move into the core product.

More access points, not more account authority

The assistant answers setup and troubleshooting questions, looks up package vulnerabilities and retrieves account information. It can also create a support case or capture a feature request. But Snyk describes access to existing account data as read-only: Assist cannot change settings, ignore issues, run scans or modify projects.

Snyk says it attaches tools to each request according to the signed-in user’s permissions. Account access is limited to the organization and group currently in use. Behind the interfaces, one LangGraph runtime—the software coordinating the agent’s work—serves a Slack app, a web app and direct API access. The team retained that runtime through each rollout phase.

250+Cases automatically detected and escalated

Snyk says Assist detected more than 250 cases and escalated them to the appropriate team.

Testing gates changes; production grades the answers

Snyk records model calls, tool calls and decision points in LangSmith, its agent-monitoring and testing system. Every proposed code change runs the real agent against evaluation suites. Changes that miss agreed thresholds are blocked. The checks cover both answer quality and attempts to make the assistant break its rules.

• Answer tests use real questions with known good answers, graded by a second model to catch deterioration after prompt or model changes.
• Adversarial tests try to make the agent ignore instructions or disclose secrets.
• Production grading checks every run for whether the question concerns Snyk and whether the response answers it.

The reported usage spans more than 500 customer accounts since the April launch. Conversations are categorized by product area, topic, programming ecosystem and error type, giving product managers a map of customer confusion and helping the company improve documentation.