Skip to content
Ayan Putatunda
Back to work

Autonomous incident response agent

production

Zendesk 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.