AWS has published a reference design for enterprise agents in which a central team maintains common controls and application teams own their software. Separate workstream cells contain runtime, memory and tool access. Deployment tests must check both allowed and forbidden requests, with human review before production promotion and procedures for rollback and retirement.
The cloud-computing provider AWS published an enterprise agent reference architecture on 28 October. It uses Amazon Bedrock and AgentCore to separate application ownership from shared platform controls. Product teams own their prompts, agent logic, tools and evaluation cases. A central team maintains common templates, versions, approved models and baseline policies.[1]
Runtime and memory are confined to individual workstream cells, each with a gateway for approved tools. A separate gateway for model requests applies model and regional permissions, quotas, request limits and budgets. Failure or rollback in one cell does not require other teams to synchronize their releases. Teams can use different development frameworks if they preserve the platform’s common contracts.[1]
The delivery pipeline first deploys a reviewed source revision outside production and evaluates the actual runtime. Allowed model calls are paired with requests to an unapproved model or region; tool and identity access receive corresponding tests. The pipeline rejects a build when a request that should have been blocked succeeds. Human review precedes production promotion, while retirement removes runtime, memory and tool grants.[1]
