Back to work
Autonomous incident response agent
productionZendesk Data Platform, 2025
Zero-touch detection, root-cause analysis, and auto-remediation of data pipeline failures using the Claude Agent SDK with Snowflake, dbt, and GitHub MCP servers.
Claude Agent SDKMCPSnowflakedbtGitHubAirflow
Problem
On-call engineers were doing the same triage every incident: read the failed run, query the warehouse, find the upstream change, open a fix. Mean time to resolution was dominated by humans reading logs at 2am.
Approach
- Agent subscribes to pipeline failures, pulls the run logs, and queries Snowflake and dbt lineage through MCP to localize the failing model.
- Correlates the failure with recent GitHub commits to the affected repos and proposes a remediation PR with the diff and its reasoning.
- Low-risk classes of failure (schema drift, late-arriving partitions) are auto-remediated; anything else pages a human with the analysis already done.
Impact
- Targeting a 70% reduction in mean time to resolution for pipeline incidents.
- On-call load shifted from log reading to reviewing a proposed fix.