Cloud cost-management vendor nOps reports in its AWS Machine Learning Blog engineering write-up that it rebuilt its FinOps assistant, Clara, on Amazon Bedrock AgentCore and cut the time needed to ship new analytics agents by 75 percent, from a 10–12 month cycle down to roughly four months.
nOps reports that it manages more than USD $4 billion in cloud spend for its customers (AWS Machine Learning Blog) and helps teams optimize commitments such as Reserved Instances and AWS Savings Plans across AWS, Google Cloud, and Microsoft Azure. Its Clara agent previously ran on a Kubernetes stack with LangChain/LangGraph orchestration and tool wrappers around web APIs — a design the team said made iteration “slow, complex, and prone to inaccuracies,” in the words of nOps engineer Jordan Stein.
The rebuild centers on a single Strands-based agent inside Bedrock AgentCore instead of a multi-agent router, which the team says avoids the latency and error propagation of agent-to-agent handoffs. Clara now runs as a Docker container defined in a single AWS Cloud Development Kit (CDK) stack, with AgentCore handling runtime, memory, guardrails, queues, and worker functions so engineers focus on domain logic rather than infrastructure.
For the analytics layer, nOps moved away from generic product APIs to governed Databricks Lakehouse Metric Views (Databricks metric views documentation), which let the agent query a predefined “True Customer Cost” measure with consistent business logic rather than recalculating pricing discounts in raw SQL each time. Durable state — sessions, canvases, and chart specs — lives in Databricks Lakebase, a serverless PostgreSQL service, so chat-generated insights can be promoted into shareable dashboards.
After the migration, nOps reports a correctness score of 81.7 percent, up 145 percent versus the prior period and roughly 65 percent in the v1 release, alongside a 79.4 percent helpfulness score, up 138 percent (AWS Machine Learning Blog). Tool failures dropped from 7.49 percent to 0.92 percent, and manual analysis that took customer-success and solutions-architecture staff about two hours now takes roughly 30 minutes — a 75 percent reduction (AWS Machine Learning Blog).
nOps says the rebuilt Clara uses an async workflow path — DynamoDB job tracking, SNS/SQS notifications, and API Gateway WebSocket push — to keep the UI updating in real time, and that for tenant isolation it runs Bedrock Guardrails as a pre-check on raw prompts while a tenant policy layer redacts internal identifiers from streamed output (AWS Machine Learning Blog).
The pattern is one more signal that enterprises are consolidating agent runtimes onto managed services like Bedrock AgentCore rather than hand-rolling orchestration. Readers building similar FinOps or analytics assistants can compare the approach with other production AgentCore deployments we have covered, such as running production AI agents in n8n with Amazon Bedrock AgentCore.
