How nOps shipped FinOps agents 75% faster with Amazon Bedrock AgentCore
Amazon Web Services Jordan Stein
nOps moved its FinOps AI to Amazon Bedrock AgentCore and says it shipped 75% faster. That cut tool failures, sped answers up, and trimmed manual analysis from 2 hours to 30 minutes.
Based on reporting by Amazon Web Services, Jordan Stein — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
nOps has rebuilt its Clara FinOps agent around Amazon Bedrock AgentCore, swapping out a layered setup built on Kubernetes, LangChain/LangGraph, and API wrappers for a more purpose-built system. The company says the change was about more than tidying up architecture. It was about making cloud cost analytics faster to ship, easier to run, and less brittle when customers ask messy real-world questions.
The old approach worked well enough to get Clara out the door, but it started to show strain as nOps scaled. API-shaped data pushed up latency and made answers less consistent. Extra orchestration and observability layers added more moving parts. And engineering time that should have gone into product work kept getting pulled into infrastructure maintenance.
The new setup centers on a single Strands-based agent running on AgentCore, with Databricks Lakehouse Metric Views handling governed analytics semantics and Databricks Lakebase storing durable state. That matters because FinOps is full of tricky definitions, from Reserved Instances to Savings Plans, and nOps wants one governed version of the truth instead of recalculating business logic every time a user asks for a metric like true customer cost.
There’s also a practical user-experience angle here. Clara now streams responses through a Vercel-hosted Next.js app, while longer jobs run asynchronously with DynamoDB job tracking and push updates through SNS, SQS, and API Gateway WebSockets. AgentCore memory keeps track of facts, user preferences, and canvas summaries, so conversations can continue across refreshes and reconnections. Guardrails and a tenant policy layer sit in front of the model to enforce cross-tenant data access rules and redact internal identifiers.
nOps says the payoff has been real. Time to production fell by 75%, from 10–12 months to 4 months. The system now serves 4 to 6 production-ready agents on a shared runtime. Correctness reached 81.7%, helpfulness hit 79.4%, and tool failures dropped from 7.49% to 0.92%. On the human side, manual analysis dropped from an estimated 2 hours to 30 minutes for Customer Success Managers and Solutions Architects. That is the kind of speedup people like to call “AI progress” when what they really mean is fewer late-night spreadsheets.
My take — AI-written commentary, not fact-checked reporting
This is the unglamorous AI story that actually matters: not a flashy model demo, but an architecture that cuts failures and gets people out of the loop. The industry keeps pretending orchestration sprawl is a feature until the bills and the bugs show up. Managed runtime, governed metrics, less glue — boring is winning again.
Read more about this at: Amazon Web Services