Ninth Wave replaced its manual, spreadsheet‑driven open‑finance onboarding process with Compass, an AI‑enabled assistant built on Amazon Bedrock AgentCore. The shift introduces a managed multi‑agent runtime that automates API validation, field mapping, and readiness scoring while preserving the security and compliance posture required for financial services.
Why a Multi‑Agent Design?
The team compared three options: self‑hosted models on Amazon EC2, a single‑agent Retrieval‑Augmented Generation (RAG) flow, and a multi‑agent architecture. Self‑hosting offered full control but added operational overhead; a single RAG agent simplified deployment but struggled with the varied mapping and analysis tasks. A multi‑agent approach, though initially more complex, allowed each specialist agent to focus on a single responsibility, keeping its context window clean and improving answer accuracy across the diverse onboarding steps.
Key Architectural Decisions
- Intent‑based routing – An orchestrator classifies the user’s request once and forwards it to the appropriate specialist agent, preventing prompt contention and ensuring predictable outputs.
- Per‑task model selection – Lightweight models handle high‑volume, low‑complexity tasks, while more capable models are reserved for mapping, analysis, and interactive Q&A, matching capability to workload.
- Tenant‑scoped grounding – Before invoking an agent, the platform injects the specific bank’s API definitions and data into the request, guaranteeing that no tenant’s data leaks into another’s session.
The orchestration logic is implemented with the Strands Agents SDK, which provides intent classification and routing between specialist agents.
Operational and Security Implications
Using Amazon Bedrock AgentCore offloads the runtime, scaling, and patching of LLM infrastructure to AWS, reducing the operational burden compared with managing EC2 instances. Per‑task model selection enables cost‑aware scaling: high‑throughput agents can run on smaller models, while complex reasoning agents can be allocated more expensive models only when needed.
Tenant‑scoped grounding aligns with multi‑tenant isolation requirements and simplifies audit logging, as each request is tied to a single bank’s context. Security controls are inherited from the broader Bedrock service stack, including IAM policies for least‑privilege access, KMS for encryption at rest and in transit, and private networking to keep traffic within the VPC. The implementation continues to satisfy Ninth Wave’s existing compliance commitments, notably SOC 2 and PCI DSS, by preserving encryption, isolation, and comprehensive logging.
Related CloudNinjas coverage: AWS.
What This Means For Practitioners
- Consider a multi‑agent pattern when a workflow spans distinct sub‑tasks that have different performance or reasoning requirements.
- Leverage managed services like
Amazon Bedrock AgentCoreto avoid the operational overhead of self‑hosting LLMs, especially in regulated environments. - Implement intent‑based routing and per‑task model selection to balance cost, latency, and accuracy.
- Ensure tenant data is injected only at the point of agent invocation to maintain strict data isolation and simplify compliance reporting.
- Map existing security controls (IAM, KMS, private networking) to the LLM runtime to meet SOC 2/PCI DSS expectations without redesigning the entire stack.