Amazon Bedrock AgentCore enables pool model multi-tenancy with full isolation

AWS has published a new architectural pattern for implementing multi-tenant AI applications using Amazon Bedrock AgentCore, focused on a 'pool model' where tenants share underlying infrastructure while maintaining complete logical isolation. The solution, demonstrated through healthcare AI agents, enforces isolation at every layer — knowledge base documents, memory, model access, and cost tracking — using native AWS capabilities. A key feature is a three-level hierarchy (Tier → Tenant → User) that enables tiered service differentiation with minimal custom code. For example, the Basic tier uses Mistral Ministral 3 8B Instruct for cost-effective document search, while the Premium tier uses OpenAI GPT OSS 120B with advanced reasoning and web search tools.

The architecture leverages Amazon Cognito for user authentication, storing tenant metadata (tier, clinic_id, role) in JWT claims that are extracted and used to scope database queries, S3 access policies, and Bedrock knowledge base filtering. This pattern is broadly applicable to SaaS platforms, enterprise solutions, and managed services. The post is part 2 of a series (with part 1 covering design considerations), and includes a GitHub repository with sample code. By combining tiering with a pool isolation model, providers can balance cost efficiency with differentiated service levels while ensuring tenant data isolation and granular observability.