nOps ships FinOps agents 75% faster with Amazon Bedrock AgentCore

nOps, an AI-powered cloud optimization solution, recently transitioned its FinOps analytics and agent experience to Amazon Bedrock AgentCore, a managed service for building and scaling agents with any framework or model. The company now better serves customers managing commitment optimization across AWS, Google Cloud Platform, and Microsoft Azure, supporting more than USD $4 billion in cloud spend under management.

Before the transition, nOps deployed Clara, its FinOps AI agent, on top of existing infrastructure including Kubernetes, Amazon Bedrock model invocation, LangChain/LangGraph orchestration, and tool wrappers around web APIs. While this allowed rapid initial delivery, it revealed structural limits: API-shaped data and large context windows increased turn latency, multiple non-purpose-built layers made iteration and debugging harder, and agent answers were tied to API responses instead of a dedicated semantic analytics layer. As Jordan Stein, an nOps leader, explained: "We were attempting to build advanced AI capabilities on top of infrastructure that wasn't designed for analytics-driven agents, which made iteration slow, complex, and prone to inaccuracies."

The new architecture centers on four layers. The interaction layer is a Vercel-hosted Next.js web application where customers interact with Clara through a conversational interface backed by streaming Server-Sent Events. The agent runtime layer runs Amazon Bedrock AgentCore with a single Strands-based agent that has direct tool access for canvas operations, query execution, datasource discovery, and workflow orchestration—avoiding the latency and error-propagation overhead of multi-agent routing. The data layer uses Databricks Lakehouse Metric Views to provide a governed semantic analytics layer; for example, a metric view called "true_customer_cost" pre-models complex business logic (amortized cost, EDP discounts, PPA credits, RI and Savings Plan components) so queries are simple and consistent across chat and dashboards. Databricks Lakebase (serverless PostgreSQL) stores durable application state including sessions, canvases, and widget specifications. An async workflow layer uses Amazon DynamoDB for job tracking, Amazon SNS and SQS for notifications, and Amazon API Gateway WebSocket push to update the UI in real time during long-running analytics tasks.

Claude runs as a Strands-based agent that supports streaming responses and tool calling; users can invoke the same workflows manually that Clara calls through Strands tools. Streaming is central to the experience: a custom merge layer sits between the Strands async stream and Server-Sent Event output, handling heartbeats to keep connections alive, text buffering with word-boundary-aware flushing to prevent UI flicker, and interleaving real-time canvas update events into the stream. Clara uses AgentCore memory with three strategies—semantic facts (organizational context like account structures), user preferences (layout choices, default aggregations, chart types), and canvas summaries (analytical thread across sessions)—so it learns how each user engages with their data over time. Sessions are scoped by canvas, not ephemeral HTTP session, so conversation context persists across browser refreshes. For tenant isolation, Amazon Bedrock Guardrails runs as a pre-check on raw user prompts, enforcing cross-tenant data access policies and prompt-attack detection, while a tenant policy layer sanitizes outbound stream chunks and widget events.

The outcome has been dramatic. Development velocity jumped: time-to-production dropped from 10–12 months to 4 months after replacing the self-managed Amazon EKS stack with a single managed service. The new runtime now serves 4–6 production-ready agents across a shared architecture. By restructuring Clara around Amazon Bedrock AgentCore and a Lakehouse-first analytics model, nOps established a more scalable foundation for AI-powered FinOps workflows while freeing engineering resources for product innovation rather than infrastructure maintenance.