Agent in a Box

Autonomous Cloud FinOps & Infrastructure Optimization Agent

engineering

Autonomous Cloud FinOps Agent: Infrastructure Cost Optimization

1. Problem Statement

Startups scaling on AWS, GCP, or Azure frequently suffer from "Cloud Sprawl"—a phenomenon where infrastructure costs grow exponentially faster than user acquisition. Engineering teams, focused on shipping features, rarely have the bandwidth to perform deep-cost inspections. This leads to three specific financial drains: unattached resources (orphaned EBS volumes, idle ELBs), over-provisioned instances (running a 16GB RAM instance for a process that uses 2GB), and missed opportunities for Reserved Instances (RIs) or Savings Plans.

Current solutions are either passive dashboards (CloudHealth, AWS Cost Explorer) that require manual intervention or "black-box" automation that engineers don't trust to touch production environments. There is a lack of an "Agentic" middle ground: a system that can analyze utilization patterns, cross-reference them with business cycles (e.g., lower traffic on weekends), and present "Pull Request-ready" infrastructure changes. For a Series A startup spending $20k/month on cloud, inefficiencies typically account for 25-30% of the bill. Much like an Autonomous R&D Tax Credit Compliance Agent optimizes tax recovery, this agent automates the detection, right-sizing logic, and notification workflow, ensuring the infrastructure footprint is lean without compromising system reliability.

2. What the Agent Does/Doesn't Do

What it does:

  • Scans cloud billing exports and real-time utilization metrics (CPU, RAM, IOPS).
  • Identifies "zombie" resources (unattached disks, old snapshots).
  • Recommends instance type changes based on historical 95th percentile usage.
  • Calculates the ROI of switching to Spot Instances for non-critical dev/staging environments.
  • Drafts Terraform/OpenTofu code snippets to implement the suggested changes.

What it doesn't do:

  • It does not execute changes directly in production without human approval (no "auto-terminate").
  • It does not manage application-level code optimizations (e.g., refactoring SQL queries).
  • It does not handle physical hardware or on-premise data centers.

3. Workflow

  1. Ingestion: Agent pulls daily Cost Usage Reports (CUR) and CloudWatch/Datadog metrics via API.
  2. Anomaly Detection: Identifies spikes in spending or resources with <5% utilization over a 7-day rolling window.
  3. Optimization Logic: Cross-references idle resources against the codebase (Terraform/Pulumi) to find where they are defined.
  4. Proposal Generation: Generates a structured report detailing: Current Cost, Proposed Change, Estimated Savings, and Risk Level (Low/Med/High).
  5. Infrastructure-as-Code (IaC) Drafting: Creates a Git branch with the modified .tf files reflecting the right-sized instance types.
  6. Notification: Pings the #devops Slack channel with a summary and a link to the Pull Request.

4. Tool Stack

  • Infrastructure Monitoring: AWS CloudWatch / Vantage.sh (Free tier / $30/mo)
  • LLM Orchestration: LangChain or CrewAI running on Claude 3.5 Sonnet (Superior reasoning for technical logs).
  • Infrastructure as Code: Terraform / GitHub Actions.
  • Cost Analysis: Infracost (Cloud Pricing API - Free for Open Source/Startup tier).
  • Pricing: Estimated total cost $50-$150/mo depending on cloud spend volume.

5. Prompt Skeletons

### Prompt 1: Utilization Analyst
You are a Cloud FinOps Expert. Analyze the following JSON metrics for an AWS EC2 instance:
Metrics: {{utilization_data}}
Current Instance Type: {{instance_type}}
Region: {{region}}

Identify if this instance is over-provisioned. Use the 95th percentile of CPU and Memory usage. 
If usage is consistently below 20%, suggest the next smallest instance within the same family (e.g., t3.medium to t3.small).
Calculate the monthly savings based on the AWS Public Price List.
Output a JSON object with: {instance_id, current_cost, proposed_instance, savings, risk_factor}.

### Prompt 2: IaC Architect
You are a DevOps Engineer. I need to modify a Terraform file to reflect a right-sizing recommendation.
Original Code:
{{original_terraform_code}}

Recommendation: Change instance_type from {{old_type}} to {{new_type}}.
Keep all other tags and configurations intact. 
Provide only the updated resource block in HCL format.

6. Success Metrics

  • Cloud Bill Reduction: Target 15-25% reduction in monthly spend within 60 days.
  • Waste Detection Rate: % of orphaned resources identified vs. total cloud footprint.
  • Engineering Friction: Hours spent by engineers on manual cost-review (Target: <1 hour/month).
  • Accuracy: % of agent recommendations accepted by the DevOps team without modification.