Home: Iago Bonifácio

Second Nature, proof of concept

An AI agent that audits the data platform

Proof of concept of a read-only agent that reviews platform telemetry and, unprompted, found a leaked key and a pipeline failing about 40% of the time with no alert.

A scheduled run starts a read-only agent that queries BigQuery, Datadog, GitHub and GCP over MCP and posts findings to Jira and Slack.Scheduleaudit runAI agentread onlyJira, SlackfindingsMCPBigQuery, Datadog, GitHub, GCP
A scheduled run starts a read-only agent that queries BigQuery, Datadog, GitHub and GCP over MCP and posts findings to Jira and Slack.

Context

On a data platform, pipelines fail silently, credentials leak, Terraform drifts from what is deployed, and cost climbs without anyone noticing. Checking all of that by hand does not scale.

Problem

Get a review with a senior engineer’s judgment, with no dedicated server and without giving an autonomous agent write access to production.

What I did

  • Read-only agents on Claude’s managed agent infrastructure, connected over MCP to Jira, Slack, GitHub, BigQuery, Datadog and the GCP APIs.
  • Scheduled audit routines and a small set of reusable skills instead of one monolithic script.
  • Ran it against the real platform as a proof of concept.

Result

The proof of concept found real issues nobody had seen: a leaked GCP service account key, a pipeline with about 40% partial failures and no alert, permission errors in staging repeated hundreds of times, run drift in Terraform Cloud, and a cost hotspot in BigQuery.