AWS published an Amazon Bedrock AgentCore reference architecture that separates central platform controls from isolated team environments and evaluates agents before production release.
On October 28, 2026, AWS published a reference architecture for organizations building and operating AI agents across multiple teams. Built on Amazon Bedrock and Amazon Bedrock AgentCore, the blueprint divides responsibilities between a central platform team and product or domain teams that develop individual agents.
The design gives each workstream an isolated account boundary containing its own agent runtime, memory, approved tools and deployment pipeline. A central control plane manages approved models, shared inference gateways, policy and reusable templates, while a separate gateway restricts which tools each agent can access. AWS says this federated setup is intended to let teams release and operate agents independently without abandoning shared governance.
The proposed release process puts agents through a series of checks before production promotion. After code review, builds and security checks, the pipeline deploys an agent to a nonproduction environment and tests the running system. Evaluations cover quality, regressions, tool performance, refusal behavior, first-token latency and cost. The tests also check that prohibited model, tool or cross-account requests are rejected; a missing expected denial stops the build. Human review precedes production release, and the blueprint includes rollback and resource-retirement procedures.
AWS lists reference templates for Strands, LangGraph and CrewAI, while allowing other frameworks if they preserve the same security and delivery controls. The architecture also describes centralized logs and metrics, per-team budgets and cost attribution, and short-lived credentials. AWS says its blueprint’s deployment and available features are based on the eu-west-1 (Ireland) Region; other regions or broader rollouts require their own validation.
The post is a technical reference rather than a claim that one architecture fits every organization. AWS recommends proving onboarding, promotion, rollback, observability and complete teardown with one representative team environment before replicating the setup.