ZeroDrift: Autonomous L2 SRE

Inspiration

The current state of Cloud Operations and FinOps is fundamentally broken. Observability tools and billing dashboards are excellent at one thing: creating alert fatigue. When an infrastructure drift occurs or a cluster is over-provisioned, dashboards merely shout at an exhausted Level-2 SRE to fix it manually. We realized that bridging the gap between detecting an anomaly and remediating it shouldn't require a human reading an alert and writing Terraform. We wanted to build a system that shifts cloud management from reactive monitoring to autonomous, self-healing execution.

What it does

ZeroDrift is an autonomous Level-2 Site Reliability Engineer. It doesn't just answer questions; it acts. Instead of sending a Slack ping about an oversized EC2 instance or a bloated Kubernetes node, ZeroDrift takes full ownership of the remediation loop:

1. Ingests & Analyzes: It reads cloud usage and billing metrics to identify idle or over-provisioned resources.
2. Reasons: It determines the exact scaling down required without impacting application availability.
3. Acts (The GitOps Way): It securely connects to the infrastructure repository, checks out a new branch, edits the Terraform/YAML state files, and opens a formatted Merge Request (MR). The SRE's job is reduced to a single click: Review and Approve.

How we built it

We engineered ZeroDrift by treating the AI not as a chatbot, but as an orchestration engine, utilizing a cutting-edge Agentic stack:

• The Reasoning Core: We utilized Google Cloud Agent Builder and Gemini 3 for advanced multi-step planning and contextual reasoning over complex infrastructure configurations.
• The Actuator (GitLab MCP): To give the agent read/write capabilities, we integrated the GitLab MCP server. This allowed Gemini to safely clone, parse, and mutate Infrastructure-as-Code (IaC) without breaking standard CI/CD pipelines.
• The Context (Fivetran MCP): We conceptually routed billing and utilization data pipelines to give the agent real-time ground truth about cloud spend.
• AI Observability (Arize Phoenix MCP): We integrated production-grade tracing to monitor the agent's decision-making process. Every MR generated by ZeroDrift has an auditable trace, proving why the AI decided to downsize a specific resource.

Challenges we ran into

Moving an LLM from a "text generator" to a "Terraform editor" presented severe blast-radius concerns. The primary challenge was preventing hallucinations inside configuration files—a single bad YAML indentation can take down production. We overcame this by enforcing strict prompt constraints, utilizing Google Cloud's Safety Settings, and architecting a strict "Merge Request Only" boundary. The agent is deliberately sandboxed; it can propose code, but it cannot merge to main autonomously.

Accomplishments that we're proud of

We successfully crossed the chasm between generative AI and deterministic GitOps. Seeing our agent autonomously read a Terraform state, understand that a t3.xlarge should actually be a t3.medium based on usage data, and push a flawless, syntactically correct MR via the GitLab MCP was a massive breakthrough. We built a tool that is genuinely enterprise-ready and safe to deploy in Day-2 operations.

What we learned

We learned that the true power of Agentic AI lies in the Model Context Protocol (MCP). By decoupling the reasoning engine (Gemini) from the tool execution (GitLab/Arize), we could build a highly modular and secure system. We also learned that building AI for SREs isn't about replacing engineers; it's about elevating them from "YAML editors" to "system approvers."

What's next for ZeroDrift: Autonomous L2 SRE

FinOps is just the beginning. Our next phase is integrating Elastic MCP to ingest real-time application logs. We aim to evolve ZeroDrift into a fully autonomous Incident Response agent that can detect a P1 outage, identify the problematic recent deployment, and automatically generate a rollback Merge Request within seconds.

Built With

• agent-builder
• arize
• devops
• fivetran
• gemini
• gitlab
• gitops
• google-cloud
• python
• sre
• terraform