AWS Launches Agent Registry to Tame Enterprise AI Agent Sprawl

AWS has launched Agent Registry, a new service inside Bedrock AgentCore that gives enterprises a single governed catalog for AI agents, tools, and skills across their organization. The preview went live on April 14 in five AWS regions — US West (Oregon), US East (N. Virginia), Europe (Ireland), Asia Pacific (Tokyo), and Asia Pacific (Sydney).

The pitch is simple: large companies have built dozens — in some cases hundreds — of AI agents across teams, and nobody can keep track of them. Agent Registry is AWS's attempt to make that mess manageable, with a private catalog, semantic and keyword search, approval workflows, and CloudTrail audit trails.

Context: The Agent Sprawl Problem

The last 18 months of enterprise AI have produced a predictable problem. Every business unit is building agents on whatever framework shipped first — LangChain, CrewAI, Bedrock Agents, AutoGen, direct API integrations against Claude and GPT. No registry, no deduplication, no shared governance. Platform teams often do not know an agent exists until it causes an incident.

This is the next phase of every enterprise technology wave. Microservices had service meshes. APIs had gateways. Agents now have Agent Registry.

Why this matters: enterprises are starting to treat agents as a managed asset class rather than a collection of one-off projects. That shift usually precedes real scale — because nobody deploys 1,000 agents without a catalog, and nobody runs them in regulated environments without audit trails.

The Details

Agent Registry is a private catalog and discovery layer for five types of resources: agents, tools, skills, MCP (Model Context Protocol) servers, and custom resources. Organizations publish resources to the registry, tag them with metadata, and control who can discover or invoke what through AWS's standard IAM permission model.

Three capabilities stand out. The first is semantic search: builders can query the registry in natural language — "find an agent that handles refund requests in Spanish" — and get ranked results based on embeddings, not just keyword matches. The second is approval workflows, which allow platform teams to gate what becomes discoverable org-wide, separating experimental agents from production ones. The third is CloudTrail integration, which logs every publish, discover, and invoke event for audit.

Access comes in three forms. There is a console UI for browsing. There is a standard AWS SDK and CLI interface for programmatic access. And — most interesting for developers — the registry is exposed as an MCP server itself, meaning agents and IDEs that speak MCP can query the registry and invoke registered resources directly. That last detail quietly locks Agent Registry into the emerging MCP ecosystem rather than positioning it as an AWS-only standard.

Industry Impact

For enterprise platform teams, this is the kind of infrastructure launch that immediately changes procurement discussions. Organizations currently debating whether to build their own agent governance layer now have a first-party option, which typically slows internal projects and accelerates vendor consolidation.

For competing cloud providers, Agent Registry raises the bar on what "AI platform" means. Microsoft has been pushing Copilot Studio and Foundry; Google has Vertex AI Agent Builder. Neither currently offers a directly comparable registry-as-MCP-server experience, and both will likely need a response. Expect announcements from at least one of the two within the quarter.

For agent framework startups — LangChain, CrewAI, and the dozen smaller players — this is more complicated. Agent Registry is framework-neutral by design and welcomes agents built on third-party stacks. But once the registry becomes the system of record inside an AWS-native company, the gravity pulls toward AWS-native tooling over time. Framework startups will need to argue they are the best place to build agents even when the registry lives somewhere else.

Expert Perspective

Enterprise analysts have been consistent about what they want from agent platforms: governance first, capabilities second. The common complaint in CIO surveys over the last two quarters has not been "AI agents cannot do enough" — it has been "we do not know what our AI agents are doing."

Agent Registry's framing directly addresses that concern. The product language emphasizes visibility, control, and reuse — three words that appear in almost every enterprise AI governance RFP. The fact that AWS shipped the feature under the AgentCore brand, rather than as a separate product, also signals the company's intent to make governance table stakes for its agent platform rather than an add-on.

The open question is one of scope. The preview is limited to five regions and does not yet cover every integration a large enterprise would want — no native hooks into Salesforce agents, ServiceNow workflows, or Microsoft 365 Copilot extensions. Expect those to arrive over the next several months as AWS expands the partner surface.

What's Next

For teams inside AWS, the immediate step is to register existing agents during the preview, even if only for inventory purposes. The product gets more valuable with more content in it, and early registrations tend to become the canonical entries when the service goes GA.

For teams on other clouds, watch the Microsoft and Google responses. If both ship equivalent registries within the quarter, the market settles into a three-way split. If neither ships, AWS accelerates its lead on the governance layer — which historically outlasts specific models and frameworks.

Pricing will also shape adoption. AWS has not detailed billing for Agent Registry yet, and the cost model it picks will determine whether enterprises build large catalogs or heavily consolidated ones.

Bottom Line

Agent Registry is not the kind of launch that trends on X, but it is the kind that quietly reshapes how enterprises run AI. The sprawl problem is real, the demand for governance is real, and AWS is the first hyperscaler to ship a first-party catalog built on open standards like MCP. Expect fast follow-ups from competitors and a material impact on how enterprise agents are deployed through the rest of the year.