AWS’s Bedrock AgentCore architecture is a federated model: a central Platform control plane governs shared services, while product or domain teams run agents, tools and data in isolated Workstream cells. The split gives teams room to own their work within shared rules for delivery, access and operations.

Shared governance, team-owned cells

A control plane is the layer that sets and coordinates rules for a system. In AWS’s design, the Platform team manages versioned templates, governance records in AWS Agent Registry, shared inference gateways, approved models and Guardrails, reviewed releases, and signals across the fleet. Guardrails are controls applied to model use.

Workstream teams retain ownership of their agents, tools and data. Their Workstream cell is an isolated environment for that team’s workload and resources, and the unit AWS uses to scale the design. Each cell includes AgentCore Runtime to execute agents, actor-scoped AgentCore Memory, and AgentCore Gateway to expose approved tool targets. “Actor-scoped” means the memory is associated with an actor. The architecture also assigns delivery and operational resources to the team.

AWS groups the design into four planes: developer experience, platform control, Workstream execution, and assurance and operations. The enterprise blueprint adds contracts for ownership, account boundaries, software delivery, observability, cost governance and retirement around AgentCore’s modular services. Its stated cost controls include quotas and budgets, tagging, chargeback, AWS Cost Explorer and CUR reconciliation.

For organizations with several agent teams, the practical distinction is who owns what: the Platform team sets shared controls, while each Workstream team owns its cell. AWS names reference templates for Strands, LangGraph and CrewAI. Teams can use a different agent framework if its template preserves the platform’s inference, tool, identity, policy, evaluation and delivery contracts.

Model calls and tool access have separate controls

AWS describes separate governed routes for model inference and tool use. Inference means sending a request to a model. Its gateway applies approved model and Region lists, Guardrails, quotas and budgets, along with identity and attribution controls.

For tools, AgentCore Gateway uses authenticated access through the Model Context Protocol (MCP), a protocol for discovering and invoking tools. The design restricts access to approved targets and uses least-privilege roles, which grant only the permissions needed for the task. Separating these routes lets the platform apply different controls to model calls and actions through tools.

How AWS gates a release to production

AWS describes a seven-stage release sequence. It evaluates the deployed Runtime, pairs authorized success tests with adversarial attempts at relevant boundaries, and requires human approval before production promotion.

StageRelease stepPurpose or gate
1Review a pull request and pin the intended immutable source revision.Establish the reviewed version to build.
2Run checks and synthesize the build, manifest, scanned image, policies, Registry descriptors and generated infrastructure.Inspect build outputs and security findings.
3Establish Workstream roles, grant exact tool aliases through the Platform pipeline and verify live policies.Confirm role and policy setup before deployment.
4Deploy Runtime, Memory, Gateway, observability, alarms, budgets, tags and application resources to nonproduction.Prepare the environment for evaluation.
5Evaluate the deployed Runtime with authorized calls and adversarial tests of model, tool, identity and throttling boundaries.A missing expected denial fails the build.
6Have a person review the deployed evidence, limitations and residual risk.Require human approval before production.
7Promote the reviewed revision, observe it, then check update, cancellation, rollback, restore and rerun recovery paths.Test the release and its recovery procedures.

The gate tests more than whether an agent can complete an allowed task: it also checks whether the deployed system denies disallowed model, tool or identity use. AWS’s sequence places that evaluation before human review and production promotion.

Retirement is part of the design too. AWS describes revoking Platform alias grants first, then removing the Workstream Gateway and tool roles, Runtime, Memory and application resources, followed by a check for residual resources.

Regional scope and the Europe (Spain) example

AWS says the blueprint’s reference deployment and available features are based on eu-west-1 (Ireland). That scope describes the blueprint, not the regional availability of every AgentCore service.

In an AWS feature-by-Region matrix snapshot inspected on October 30, 2026, AgentCore Memory and AWS Agent Registry were marked unsupported in Europe (Spain). Both appear in the blueprint, so teams in that Region cannot assume the reference design will deploy unchanged. The matrix describes feature availability by Region; it does not establish that the complete blueprint has been validated there.