Autonomous Reliability & Operations Agent (AROA)

Detect, diagnose, and remediate data pipeline incidents using Fivetran, Elastic, and MongoDB

Comment

Inspiration

Most data teams still discover broken pipelines when a dashboard looks wrong or a stakeholder pings them on Slack. By then, it’s already too late. We wanted an agent that behaves like a Data SRE for Fivetran pipelines: it should watch syncs, understand lineage and impact, and propose safe fixes instead of just throwing error messages.

The Google Cloud Rapid Agent Hackathon pushed us to go beyond chatbots and build a real agent that reasons, plans, and calls tools using Gemini and a partner MCP—in our case, Fivetran, with Elastic and MongoDB as supporting cast.

What it does

Autonomous Reliability & Operations Agent (AROA) is a web console plus Gemini‑powered agent that sits on top of your Fivetran‑style pipelines and observability stack. In the demo environment, it:

Shows a Connector Health Dashboard 2.0 with per‑connector sync history, duration trends, row counts, and status (active, failed, paused, warning).

Tracks schema changes and downstream impact: when a connector’s tables or columns change, AROA shows a timeline and which dashboards/views are affected.

Provides a Cost & Usage view (MAR‑style) with synthetic usage and credit estimates, helping prioritize which connectors need optimization.

Correlates connector status with Elastic‑style logs and BigQuery‑style freshness, then updates MongoDB‑style incident records with severity, status, and playbooks.

Exposes an AI‑assisted “Add connector” flow: you describe a new pipeline in natural language and Gemini proposes a connector config that can be added to the demo environment in one click.

Even though this hackathon version uses curated demo data, the endpoints, schemas, and agent tools are designed so you can plug in real Fivetran, Elastic, MongoDB, and BigQuery instances with minimal changes.

How we built it

We started in Google AI Studio / Agent Builder, defining a system prompt for a Data Reliability SRE and tool schemas that mirror real operations on top of:

Fivetran‑style connectors (list, status, logs, schema changes, sync history, usage).

Elastic‑style logs (search, error rate, correlated errors).

MongoDB‑style incidents (create/list/update incidents, fetch playbooks).

BigQuery‑style freshness & impact checks (latest timestamps, downstream views).

On the backend, we used Node.js + TypeScript + Express to implement /api/fivetran, /api/elastic, /api/mongo, /api/bigquery, plus /api/chat as a bridge to the Gemini Interactions API with our managed Agent ID. The Fivetran service returns rich demo objects shaped like Platform Connector / Metadata API outputs—sync runs, schema events, and usage metrics—so the agent has something realistic to think with.

On the frontend, we built a React + Vite single-page app with routes for Overview, Pipelines, Observability, Incidents, Settings, and the AI‑assisted connector flow. The Pipelines page now has tabs for Health, Schema & Impact, Logs, and Usage. The Incidents page is backed by an in‑memory MongoDB‑style store with filters, status transitions, and “Ask the agent for a remediation plan” integrated via /api/chat.

Challenges we ran into

Tool schemas and validation: Getting Gemini, AI Studio, and our backend to agree on JSON schemas required a lot of iteration—especially around optional fields, required arrays, and staying within the supported subset of JSON Schema for tools.

Designing realistic Fivetran demo data: We wanted the demo to feel like a real Platform Connector / Metadata feed—sync histories, MAR‑style usage, schema change events, and lineage—without having access to live customer data. Balancing realism with simplicity was tricky.

Coordinating multiple systems through one agent: Making the agent reason coherently across Fivetran health, Elastic logs, Mongo incidents, and BigQuery freshness demanded careful tool design and prompt engineering so it wouldn’t over‑orchestrate or miss obvious steps.

Accomplishments that we're proud of

Turning Fivetran from “a couple of status flags” into a full reliability surface: per‑run history, schema changes, usage, and impact are all visible and actionable from one console.

Building an AI‑assisted connector designer that leverages Gemini to translate a plain‑English description like “Sync Shopify orders into BigQuery hourly” into a structured connector proposal.

Making the agent multi-modal in behavior (not just chat): it scans connectors, opens incidents, consults playbooks, and calls simulated schema/usage APIs before presenting a remediation plan.

Keeping the architecture production‑shaped: swapping out the demo stores for real Fivetran REST/MCP, Elastic Cloud, MongoDB Atlas, and BigQuery would be straightforward.

What we learned

Fivetran’s metadata is a goldmine: once you model sync runs, schema events, and usage in a consistent way, reliability and cost insights fall out naturally.

Good agents depend on strongly scoped tools; forcing every action through clear, side‑effect‑aware functions dramatically improved Gemini’s planning behavior.

Even in a hackathon setting, investing in UX for SRE‑style workflows (health dashboards, schema timelines, incident timelines, approvals) pays off more than adding one more clever prompt.

Integrating Gemini via the Interactions API and Agent Builder gave us a solid pattern for future production agents—server‑side history, proper tool calling, and a clean separation between reasoning and UI.

What's next for Autonomous Reliability & Operations Agent (AROA)

Post‑hackathon, we’d like to:

Connect to real Fivetran accounts using the REST API or MCP server, so AROA can monitor actual connectors, not just demo data.

Use Fivetran’s Metadata API and warehouse metadata to build end‑to‑end lineage graphs, then let the agent simulate blast radius and “what‑if” scenarios when a table or connector changes.

Add deeper cost optimization recommendations, combining MAR, connector usage patterns, and downstream dashboard usage to suggest schedule changes or filtered syncs.

Integrate with incident management tools (e.g., Jira, PagerDuty) so incidents and playbooks in AROA can tie directly into existing on‑call workflows.

Expand beyond Fivetran into other partner tracks (Elastic, MongoDB, Arize, GitLab, Dynatrace) while keeping the same agent and reliability patterns.

Built With

Share this project:

Updates