Automating AWS cost cuts with CrewAI and Bedrock revealed hidden inefficiencies

AWS’s built-in optimization tools often stop short of aggressive savings. Trusted Advisor’s free tier only covers seven checks, leaving gaps like missed gp2-to-gp3 volume migrations or isolated orphaned snapshots. Compute Optimizer targets EC2 and Lambda but ignores persistent cost drains such as unattached Elastic IPs, which accumulate $3.60 per month each. Without a tool that pinpoints specific resource IDs for termination, spending insights remain too abstract.

The core limitation of AWS’s native solutions is their focus on broad health assessments rather than targeted cost reduction. They provide recommendations but lack a structured plan. To address this, I designed an autonomous workflow using CrewAI and Amazon Bedrock, transforming static dashboards into a dynamic system capable of analyzing real-time infrastructure against cost-saving logic.

By assembling a team of specialized agents—one acting as a cloud architect, the other as a financial analyst—I replicated collaborative decision-making. One agent scanned for underutilized resources while the other evaluated their financial impact. Bedrock’s reasoning layer then detected patterns standard tools overlook, such as 90-day-old snapshots with no active production ties.

The workflow prioritized overlooked cost drivers. Compute Optimizer might flag oversized instances, but it doesn’t actively hunt for unattached Elastic IPs silently draining budgets. Automating the discovery of these orphaned resources, along with gp2-to-gp3 migration opportunities, yielded a precise list of actionable resource IDs.

This method delivered a 40% monthly cost reduction, though the result wasn’t effortless. Success hinged on precise prompt engineering and clear agent constraints. Directly exposing an LLM to an AWS account without guidance won’t yield savings—it requires deliberate oversight.

The true value lies in automating granular checks. Instead of reviewing high-level spending graphs, the system generated a CSV of exact snapshot and volume IDs no longer in use. Native tools like Compute Optimizer and Trusted Advisor remain useful, but they lack the depth needed for aggressive optimization.

The combination of CrewAI’s orchestration and Bedrock’s LLM capabilities exposed hidden costs—such as $3.60 Elastic IPs and outdated gp2 volumes—enabling the 40% spending cut. The key shift was from passive monitoring to active, agent-driven waste identification.