AWS is introducing the Agent Registry in AgentCore, a centralized system for sharing and reusing AI agents within organizations. The preview version is designed to address the problem of “agent sprawl” and helps companies that deploy hundreds or thousands of agents maintain control over their AI landscape.
Companies scaling up to hundreds or thousands of agents face three challenges. First, there is a lack of visibility into which agents are running. Second, there is no control over who is allowed to publish agents. Finally, teams often build functionality that already exists.
AWS sees chaos rapidly increasing without a central system. Compliance risks rise, and developers waste time on duplicate work. The problem is exacerbated by agents being spread across AWS, other clouds, and on-premises environments.
Platform teams need more than a simple list. They must be able to build agents, publish them using workflows, help teams find existing agents, determine who can use what, monitor production, and phase out outdated agents.
What does the Agent Registry offer?
The registry stores metadata for every agent, tool, MCP server, agent skill, and custom resource. It records who published something, which protocols it uses, and how to invoke it. Standards like MCP and A2A are directly supported, with room for custom schemas.
Registration can be done in two ways. Manually via the console, SDK, or API, where you add descriptions and documentation yourself. Or automatically by referencing an MCP or A2A endpoint that retrieves the details itself.
The registry operates via the AgentCore Console, APIs, and as an MCP server. MCP-compatible clients such as Kiro and Claude Code can communicate with it directly. For organizations with their own identity providers, OAuth access enables teams to build their own discovery interfaces without IAM credentials.
Centralized Search and Reuse
Without a central registry, developers search externally or unknowingly duplicate features that colleagues have already built. The registry solves this with a hybrid search that combines keywords and semantics. Long queries also use semantic understanding to find conceptually related results.
A search for “payment processing” also shows tools tagged as “billing” or “invoicing.” Teams search the registry first before building anything new. If an approved feature exists, they use it. Otherwise, they build it, register it, and make it available to others.