Yanolja's Case Study: Building AIOps Agents with Strands SDK and Bedrock AgentCore

To reduce repetitive tasks and inefficiencies in incident response caused by increasingly complex cloud infrastructure operations, Yanolja built AIOps Agents based on Amazon Bedrock AgentCore and the Strands Agents SDK. 6 teams with 14 engineers participated in Full-day Sprints over 6 weeks, running the project in a direction that directly automated concrete pain points from the field.

The motivation was clear. There was an accumulation of issues: 8 or more standardized operational inquiries occurring daily, response delays where writing a postmortem after an incident took up to 2 weeks, complex diagnostics requiring manual analysis such as IDP failures and DB alarms, and difficulty checking AWS costs. The goal was not a simple PoC, but a 50% or greater reduction in manual work time and cultivating AI Champions within the organization.

At the core of the solution were Amazon Bedrock and Bedrock AgentCore. By leveraging enterprise-grade security, support for various Foundation Models, a fully managed operating environment, and support for 8-hour long-running workloads, the burden of Agent development and operations was reduced. In addition, a Self-correct RAG pattern was applied by combining Bedrock Knowledge Base with OpenSearch, connecting existing documents and incident history to the Agent's knowledge.

Complex collaboration between Agents was implemented with Strands Agents. In a model-driven approach where planning, delegation, and evaluation are performed based on the model's reasoning ability, an Orchestrator Agent coordinated Sub-Agents, and context sharing between Agents along with MCP integration connected to external systems such as AWS API, DataDog, GitHub, ArgoCD, and Harbor.

The 6 Agents built directly targeted each team's operational tasks.

• IDP Operations Efficiency: The Support Agent identified the intent of requests and delegated to the Knowledge Agent, Provisioning Doctor Agent, and Runtime Diagnostic Agent to automate IDP failure diagnosis and resolution.
• Slack-based Incident Response Automation: A Slack Bot and AI Agent summarize incident situations, collect related metrics and logs, and generate a postmortem draft within 24 hours.
• DB ChatBot: Using Datadog, CloudWatch, Confluence, and Knowledge Base, this automates DB alarm analysis, root cause identification, and identifying the responsible person.
• DataCenter Agent: Collects server and switch status via Ansible, and AI reflects changes in documentation to keep asset information up to date.
• FinOps Helper: Receives natural language cost queries, automatically generates Athena (CUR) SQL, and suggests cost optimization guidance.
• DD_Kitty: Calls Bedrock via Slack, SQS, and Lambda, sequentially performing GitHub Manifest analysis and configuration validation.

The results were clear as well. The 14 engineers directly experienced prompt engineering, RAG pipelines, and Agent tool integration, internalizing AI Agent development capability, and the perspective of redefining operational work as "tasks that can be delegated to an AI Agent" spread throughout the organization. They expect postmortem writing time to drop from 2 weeks to 24 hours, and MTTR to drop from 10 minutes to 5 minutes, and they set a goal of reducing manual work time by 50% or more as well.

Going forward, they plan to build a Core Agent that automatically delegates natural language requests, an Agent Builder for creating Agents without coding, multi-agent collaboration based on the A2A protocol, and a domain-specific Agent ecosystem expanding to SRE, DBA, DevOps, and Security.