AWS split agent deployment into separate managed services

The conversation

On 16 July 2025, AWS announced that "Amazon Bedrock AgentCore" was "now available in preview," describing it as a set of separate services rather than one product: AgentCore Runtime, for session isolation; AgentCore Memory, for short- and long-term memory "with zero infrastructure management"; AgentCore Gateway, which converts existing APIs "into Model Context Protocol-compatible tools"; AgentCore Browser Tool and Code Interpreter, for sandboxed web and code execution; AgentCore Observability, built on Amazon CloudWatch; and AgentCore Identity, for connecting to identity providers such as Amazon Cognito. The announcement states availability was limited at launch to four AWS regions.

What the documents show

AWS's own announcement names AgentCore's stated purpose as moving agents "from proof of concept to production," and states its services "work with any model" and "any open-source agent framework," a vendor claim about compatibility rather than an independently tested one. As retrieved 16 September 2026, the platform's FAQ page extends that claim, stating a builder can connect through "any protocol (such as MCP or A2A)," naming both protocols already on file, and "any model (such as Amazon Nova, OpenAI, or Gemini)." The current product page organizes the offering around the same seven named services as the July announcement, with case studies from named customers including Druva, whose stated resolution of "68 percent of support issues without human intervention" is that customer's own figure, attributed to Druva, not to an AWS measurement.

The system boundary

AgentCore sits beneath, not instead of, the higher-level Bedrock Agents feature already on file: both the July announcement and the current documentation describe it as infrastructure a builder assembles from separate services, in contrast to a single managed "agent" object. The boundary AWS draws for the Gateway service is protocol translation, not agent reasoning: the service turns an existing API into an MCP-compatible tool, leaving the decision of when to call that tool to whatever model or framework sits above it.

Where it fails

As retrieved on 16 September 2026, AWS's own FAQ page still describes at least some capabilities as "available in preview," including an observability feature and a stablecoin micropayments capability tied to third-party wallets, meaning the July 2025 preview label had not uniformly become a general-availability claim by the retrieval date for every part of the service, even though the same page also presents customer case studies as though in current production use.

• Which named AgentCore service does a given claim about "AgentCore" actually apply to?
• Is a capability still labelled preview in a vendor's own current documentation, and does a production deployment depend on it?
• Is a customer-reported figure, such as a resolution rate, that customer's own claim or an independently audited one?

AgentCore's own documents describe a modular, partly still-preview set of services rather than one finished product, a distinction worth preserving whenever "AgentCore" is used as a single noun.