Ninth Wave moves bank onboarding to a multi-agent system on Amazon Bedrock
Ninth Wave built Compass, an onboarding assistant running seven specialist agents on Amazon Bedrock AgentCore, to speed up open finance API integration for banks.
Banks joining open finance networks expose their APIs with their own field names and formatting conventions. Validating those APIs, mapping fields to the Financial Data Exchange (FDX) standard and scoring readiness for production has traditionally taken weeks of specialist work spread across email threads and spreadsheets, according to AWS.
Ninth Wave connects financial institutions to third-party applications, including aggregators such as Plaid, Finicity and MX and accounting systems such as Intuit QuickBooks, Xero and Sage. Its platform normalizes bank APIs to FDX, so a bank integrates once and reaches the whole network. To make onboarding self-service, the company built an assistant called Compass and runs it on Amazon Bedrock AgentCore.
Seven specialists instead of one agent
The team first tried self-hosted models on Amazon EC2, which gave full control but added operational overhead, and a single-agent Retrieval Augmented Generation (RAG) setup, which was simpler but less accurate across mapping, analysis, search and interactive Q&A. In the multi-agent design each agent handles one task with its own context and instruction prompt. The Strands Agents framework does the orchestration: it classifies intent and routes the request to one of seven specialists.
Grounding happens at the application layer. Before an agent is invoked, Compass pulls that bank's API documentation and configuration data from Amazon OpenSearch Service and Amazon S3 using per-tenant indices and prefixes. That separation is what keeps one bank's information out of another bank's session even though both share the same model infrastructure. Readiness analysis is the only specialist that retrieves from an Amazon Bedrock knowledge base, because it needs to synthesize across a corpus of FDX references.
On the security side, traffic enters through Amazon CloudFront and AWS WAF v2 with default-deny web ACLs, and sessions are validated on Amazon ECS on AWS Fargate against an OAuth2/OIDC provider with multi-factor authentication. The AI workload runs in a separate AgentCore runtime reached through a cross-account IAM role.
The full architecture is described in the AWS blog post.
