Cost Intelligence Agent: An Autonomous Agent Solution for Cost Governance and Invocation Monitoring of Amazon Bedrock

This article introduces Cost Intelligence Agent—an open-source project based on Amazon Bedrock AgentCore. It enables autonomous cost governance, invocation monitoring, and CloudWatch alerts through prompt engineering-driven workflows, helping enterprises control the costs of AI workloads. The project core addresses issues like cost tracking, anomaly detection, and root cause analysis in Bedrock usage, providing a serverless architecture and simplified deployment experience.

With the widespread adoption of Amazon Bedrock in enterprises, the cost of AI model invocations has become difficult to predict and control due to token usage (depending on input/output length). Enterprises face challenges such as: lack of fine-grained cost tracking (unable to split costs by model/agent), difficulty in timely detection of invocation anomalies, manual intervention in root cause investigation, and repeated issues due to no historical pattern learning. Cost Intelligence Agent is designed to address these problems.

Core Features: 1. Cost Governance (track spending by model/agent, enforce budgets, detect anomalies); 2. Invocation Monitoring (analyze token patterns, monitor throttling events, count invocation frequency); 3. CloudWatch Alerts (pre-configured with 5 rules, automatically initiate investigation and attach reports when triggered); 4. Prompt Engineering Workflow (hypothesis-driven investigation, evidence ledger, adaptive response).

Technical Architecture: Built on Bedrock AgentCore and Strands SDK; uses Claude Sonnet4.6 as the inference engine; Web UI with Amplify, identity authentication with Cognito; core runtime integrates 11 tools (interacts with CloudWatch, CloudTrail, etc.); event-driven data flow (CloudWatch→EventBridge→Lambda→agent investigation→notifications + DynamoDB storage); fully serverless, pay-as-you-go.

When an alert is triggered, the agent initiates a structured investigation: 1. Generate initial hypotheses (e.g., token surge may be due to agent infinite loop); 2. Collect evidence (query CloudWatch metrics, CloudTrail logs, Cost Explorer data) and record in the ledger; 3. Evaluate hypotheses (confirm/deny/revise; generate new hypotheses if evidence is insufficient); 4. Generate structured reports (summary of findings, timeline, recommended actions). It learns from historical events through pattern memory for precise root cause analysis.

Simplified Deployment: Download the CloudFormation template, run the aws cloudformation create-stack command, and complete deployment in 5 minutes (automatically creates IAM roles, ECR repositories, Lambda, DynamoDB, Cognito, and other resources). Flexible Configuration: Can specify admin email, default model (Haiku4.5/Sonnet4.5/Sonnet4.6, etc.), monthly budget cap, Slack integration (Bot Token), memory retention days, custom model ID, etc., to adapt to the needs of enterprises of different sizes.

Investigation Cost: Sonnet4.6/Sonnet4.5 costs approximately $0.25 per investigation, Haiku4.5 about $0.03 per investigation; monthly total cost depends on alert frequency and number of investigations. Infrastructure costs (alerts, DynamoDB, Lambda) are either in the free tier or negligible. Compared to manual investigation, the autonomous solution has significant economic advantages in large-scale scenarios and provides 7x24 monitoring capability.

Practical and Aesthetic Web Interface: The main dashboard displays monthly spending, budget usage rate, active alerts, and recent investigation list; the investigation details page provides a timeline view (agent operation steps, evidence, hypotheses) to enhance transparency; supports dark/light theme switching, suitable for long-term viewing.

Applicable Scenarios: Organizations where multiple teams share Bedrock resources (split costs by team/project), production environments running multiple AI agents (unified monitoring and governance), cost-sensitive workloads (timely prevention of abnormal consumption), small and medium-sized enterprises lacking dedicated operations (automated monitoring and alerts). Value: Provides a replicable AI cost governance model that will become a standard practice for enterprise AI operations.