Snyk moved its AI support agent, Snyk Assist, from an internal support tool into its core product on 1 September 2026, making it available to every paying customer. Since its customer launch in April 2026, the agent has handled more than 60,000 queries across 500+ customer accounts, resolved 85%+ of sessions without a support ticket and automatically detected and escalated 250+ cases to the right team.

Snyk Assist answers customer questions in plain language and can look up open issues, check packages for known vulnerabilities, open support cases and log feature requests. Before its launch, customers often had to search documentation, support articles, release notes, learning materials and account data, or file a ticket and wait. The agent first served Snyk’s support staff as a case triage tool and virtual agent for roughly a year, with feedback cycles measured in hours.

Snyk introduced the agent in its support portal in April 2026, where it could greet customers by name, use account context, run a login health check and raise tickets at any time. The team then placed it in a panel at the top of every page in the core product. The same LangGraph runtime serves the support portal, Slack app, web app and direct API access; Snyk says only the customer-facing surface changed across the rollout phases.

Access controls were built around each signed-in user’s permissions: tools are attached for each request, so the agent can access only data that user is allowed to see. Most tools run as separate microservices. Conversation history is stored in PostgreSQL and keyed to each session, supporting multi-turn conversations, scaling across pods and resuming sessions. Middleware provides defined points for context management, guardrails and model fallbacks; teams can add or reorder behavior with a one-line list change. A shared factory compiles the agent graph, while teams configure workflows for their needs.

Snyk traces every model call, tool call and decision in LangSmith. Its offline evaluations test questions with known answers and run automated red-team attempts to make the agent ignore instructions or reveal secrets. Every pull request runs the agent against those test suites and is blocked if it misses agreed thresholds. Scheduled evaluations grade production responses for whether they concern Snyk and whether they answer the question. The team also categorizes queries by product area, topic, language ecosystem and error type, and turns problematic traces into datasets through the LangSmith MCP server. Snyk says this process helps it catch regressions and improve documentation based on customer confusion.

Snyk moved its AI support agent, Snyk Assist, from an internal support tool into its core product on 1 September 2026, making it available to every paying customer. Since its customer launch in April 2026, the agent has handled more than 60,000 queries across 500+ customer accounts, resolved 85%+ of sessions without a support ticket and automatically detected and escalated 250+ cases to the right team.

Snyk Assist answers customer questions in plain language and can look up open issues, check packages for known vulnerabilities, open support cases and log feature requests. Before its launch, customers often had to search documentation, support articles, release notes, learning materials and account data, or file a ticket and wait. The agent first served Snyk’s support staff as a case triage tool and virtual agent for roughly a year, with feedback cycles measured in hours.

Snyk introduced the agent in its support portal in April 2026, where it could greet customers by name, use account context, run a login health check and raise tickets at any time. The team then placed it in a panel at the top of every page in the core product. The same LangGraph runtime serves the support portal, Slack app, web app and direct API access; Snyk says only the customer-facing surface changed across the rollout phases.

Access controls were built around each signed-in user’s permissions: tools are attached for each request, so the agent can access only data that user is allowed to see. Most tools run as separate microservices. Conversation history is stored in PostgreSQL and keyed to each session, supporting multi-turn conversations, scaling across pods and resuming sessions. Middleware provides defined points for context management, guardrails and model fallbacks; teams can add or reorder behavior with a one-line list change. A shared factory compiles the agent graph, while teams configure workflows for their needs.

Snyk traces every model call, tool call and decision in LangSmith. Its offline evaluations test questions with known answers and run automated red-team attempts to make the agent ignore instructions or reveal secrets. Every pull request runs the agent against those test suites and is blocked if it misses agreed thresholds. Scheduled evaluations grade production responses for whether they concern Snyk and whether they answer the question. The team also categorizes queries by product area, topic, language ecosystem and error type, and turns problematic traces into datasets through the LangSmith MCP server. Snyk says this process helps it catch regressions and improve documentation based on customer confusion.