AWS released Agent Registry in public preview this week as part of Amazon Bedrock AgentCore, creating a centralized catalog where enterprises can discover, share, and govern every AI agent, tool, MCP server, and agent skill in their organization. The differentiator: it indexes agents regardless of where they run, including on Azure, GCP, and on-premises infrastructure.
How It Works
Teams register agents in one of two ways, according to the AWS Machine Learning Blog. They can provide metadata manually through the console, SDK, or API, specifying ownership, capability descriptions, and compliance status. Or they can point to an MCP or A2A endpoint and let the registry pull details automatically.
Every record captures who published it, what protocols it implements, what it exposes, and how to invoke it. The registry supports MCP and A2A natively with the flexibility to define custom schemas.
Discovery uses hybrid search blending keyword and semantic matching. A search for “payment processing” surfaces tools tagged as “billing” or “invoicing” even when the names don’t match, per InfoQ. The registry is accessible through the AgentCore console, APIs, and as an MCP server itself, so any MCP-compatible client, including Kiro and Claude Code, can query it directly.
Governance Workflow
Records follow a draft-to-pending-to-approved lifecycle. Agents don’t become discoverable until approved. Admins control who can register and who can discover through IAM policies. Versioning tracks changes over time, and organizations can deprecate records when they’re no longer needed. Custom metadata fields let teams attach cost center, deployment environment, or security classification data.
For organizations with custom identity providers, OAuth-based access lets teams build their own discovery interfaces without requiring IAM credentials, per the AWS blog.
Early Adopters and Rough Edges
Southwest Airlines is an early user. Justin Bundick, VP of AI and Intelligent Platforms, said the registry “addresses the critical discoverability challenge, enabling teams to find and reuse existing agents rather than rebuild capabilities from scratch,” per the AWS announcement.
Zuora, which deploys 50 agents across sales, finance, product, and developer teams, is also onboard. Chief Product and Technology Officer Pete Hirsch said the registry gives principal architects a unified view for discovering and cataloging every agent in use.
Solutions architect Shinya Tahara ran hands-on testing and found some preview-stage limitations, as reported by InfoQ. Semantic search worked well with English queries but stumbled on Japanese queries against English-only metadata, returning zero results for one of three Japanese test queries. Folding filter constraints into natural language queries degraded precision significantly, returning all registered records instead of targeted results. And updating any record resets its status to DRAFT, requiring re-approval, a friction point for teams managing frequently updated agents.
The Governance Infrastructure Arms Race
Agent Registry shipped in the same week as Databricks Unity AI Gateway (which extends data governance to agentic AI with MCP governance and per-request cost logging), Microsoft Foundry Toolkit (with MCP tool approval controls), and Salesforce AgentExchange (a unified marketplace with governance for 13,000+ listings). Every major cloud and platform vendor is now shipping agent governance infrastructure simultaneously.
The Cloud Security Alliance published a study this week finding that 54% of organizations report unsanctioned shadow AI agents in their environments. Agent Registry is AWS’s direct answer to that inventory problem: a catalog that can see agents across every cloud and on-premises deployment, with approval gates before any agent becomes discoverable to the broader organization.