AWS launched Agent Registry in preview on April 9 through Amazon Bedrock AgentCore, and it is the enterprise-governance layer that the industry has been pricing in since the "AI sprawl" data started flowing this quarter. The registry is a private, governed catalog for agents, tools, skills, MCP servers, and custom resources inside an organization. Access is through the AgentCore Console UI, AWS CLI and SDK, or as an MCP server that developers query and invoke directly from their IDEs. Features include semantic and keyword search, approval workflows, CloudTrail audit trails, and URL-based auto-discovery that pulls tool schemas from live MCP endpoints. Preview availability is in five regions: US East (N. Virginia), US West (Oregon), Asia Pacific (Sydney), Asia Pacific (Tokyo), and Europe (Ireland). AWS is positioning the registry against three specific problems: visibility gap in what agents exist, weak governance over deployment, and tool duplication across teams.
The design choices are worth reading carefully. First, the registry is itself queryable as an MCP server. That is the same pattern OpenAI adopted in its Agents SDK update yesterday, and it confirms MCP as the governance-interface standard, not just the agent-tool-call standard. Second, URL-based auto-discovery means the registry can ingest an existing MCP endpoint and extract its tool schema automatically; that reduces the per-tool registration overhead that kills most internal registry projects at enterprise scale. Third, approval workflows plus CloudTrail audit trails provide the compliance and security posture that enterprise buyers require before they allow agent sprawl to continue; AWS is pitching this as the "control layer" for an agentic footprint that has already outgrown what most organizations can govern by hand. Fourth, the five-region scope is deliberate: the chosen regions cover most regulated-industry and enterprise deployments, and notably exclude China and India for geopolitical and regulatory reasons.
This validates the 2026 thesis for enterprise AI spend that has run through several pieces covered here. The Grant Thornton "AI proof gap" (97 percent of executives report benefits but only 29 percent see organizational ROI), the OutSystems 94 percent AI-sprawl concern, and the Gartner 40-percent-cancellation forecast by 2027 all point at the same remediation category: visibility, governance, approval, audit. AWS shipping Agent Registry inside Bedrock AgentCore is the first major-hyperscaler answer to that demand. AWS is also doing it at the same time it ships Claude Mythos preview access through Bedrock (April 13), which stacks the hyperscaler's two 2026 agent-AI bets in the same platform: access to the gated frontier models, plus the governance layer to manage them responsibly. Azure and Google Cloud will ship equivalent registries within two quarters, probably with more MCP-first positioning. The consultancies selling AI-remediation engagements still have a window, but the platform players are closing the obvious gap first.
For AWS-committed enterprises, Agent Registry is worth piloting immediately if your agent footprint has outgrown informal coordination; that usually happens at the twenty-to-fifty-agent threshold, and the CloudTrail integration makes this the easiest compliance-friendly registry to stand up inside an already-AWS stack. For multi-cloud and open-source-first teams, the interesting signal is that AWS is being deliberately cloud-agnostic in messaging: the registry supports MCP endpoints regardless of where they run, which suggests AWS sees agent sprawl as a horizontal problem and is trying to position the registry as the control point even for non-AWS agents. For independent tooling builders, the competitive window for open-source agent registries narrowed today; you either differentiate on something AWS cannot credibly do (deep multi-cloud, self-hosted, niche compliance postures) or get acquired or absorbed. For everyone, the signal to internalize is that agent governance has crossed from "interesting future problem" to "in-preview product category at the major hyperscaler," and planning against that reality is now a 2026 item, not a 2027 one.