An AI agent that audits its own data platform

A Claude-native automation layer that reads platform telemetry and surfaces real production risk — leaked credentials, silent pipeline failures, cost drift — without a human triggering the check.

• Claude (Managed Agents)
• MCP
• Jira
• Slack
• GitHub
• BigQuery
• Datadog
• GCP

Context

Running a production data platform means a constant background risk: pipelines fail silently, credentials leak, Terraform drifts from what’s actually deployed, and cost creeps up unnoticed. Catching all of that by hand doesn’t scale.

Problem

Building an automation layer that watches the platform the way a senior engineer would — without standing up and maintaining a dedicated server, and without giving an autonomous agent write access to production.

Approach

• Designed a read-only agent automation system on top of Claude’s managed agent infrastructure, wired to the platform’s own tools via MCP (Jira, Slack, GitHub, BigQuery, Datadog, GCP APIs).
• Shipped scheduled, read-only audit workflows plus a small set of reusable skills, rather than one monolithic script.
• Ran it against the real platform as a proof of concept.

Outcome

The proof of concept found real, previously-unnoticed issues on its own: a leaked GCP service-account key, a pipeline with a roughly 40% partial-failure rate that wasn’t alerting, staging permission errors recurring hundreds of times, Terraform Cloud run drift, and a cost hotspot in BigQuery — all surfaced without a human going looking for them first.

Lessons

Read-only-by-design plus a small catalog of focused workflows beats one autonomous agent with broad write access — it found real issues without ever needing production write permissions.