AWS Agent Registry: How a Private Catalog Prevents AI Agent Chaos

The early days of AI adoption often follow a predictable script. A team builds an agent to solve a problem. Another team, unaware of the first effort, builds the same tool. Months later, the CTO realizes they’ve accidentally created a parallel IT infrastructure—one that burns budgets, duplicates workflows, and introduces security gaps that no single source of truth can resolve.

This scenario isn’t hypothetical. It’s playing out in companies across Latin America and beyond, where platform leads describe a growing nightmare: uncoordinated AI agents multiplying like wildfire, each with its own guardrails, APIs, and maintenance backlogs. The critical question isn’t whether these tools are useful—it’s whether organizations can govern them before chaos sets in.

That’s where AWS’s new Agent Registry comes in. Announced on April 9, 2026, this private catalog is designed to bring order to the AI agent sprawl by providing a single place to publish, validate, and approve every agent, MCP server, skill, or custom resource in an organization. It’s not a build tool—it’s a discovery and governance service that forces early discipline into a process that, if left unchecked, could easily spiral into a migration nightmare.

The Hidden Costs of Unmanaged AI Agents

The consequences of agent sprawl extend far beyond duplicated code. Consider these real-world examples from organizations grappling with their first wave of AI deployments:

• Budget drain: Two separate teams build agents that call the same Jira API, each consuming Bedrock tokens and racking up costs without oversight.
• Security gaps: A billing team’s MCP server for Stripe internally reuses a PII redaction function that the Compliance team already built as a standalone tool—creating redundant logic and potential vulnerabilities.
• Workforce inefficiency: The Analytics team spends weeks rebuilding an Athena wrapper that the Data team had already created, simply because no one knew it existed.
• Operational confusion: When a production incident occurs, teams scramble to determine which agent is the "official" one, leading to delayed resolutions and duplicated efforts.

These aren’t isolated incidents. They’re symptoms of a systemic problem: agent governance isn’t an afterthought—it’s a day-one requirement. The difference between a smooth deployment and a costly retroactive cleanup is measured in weeks, if not months, of lost productivity.

How AWS Agent Registry Works

AWS Agent Registry isn’t just another catalog—it’s a structured system designed to enforce governance from the start. Here’s how it functions in practice:

What Gets Cataloged

The Registry supports four types of records, each validated against specific schemas to ensure consistency:

• MCP servers: Tools that expose capabilities to agents, validated against the official MCP schema.
• Agents: Corporate agents that other agents can invoke, validated against the A2A AgentCard schema.
• Skills: Reusable capabilities like Python packages or libraries, accompanied by metadata and documentation.
• Custom resources: Flexible JSON-based entries for anything that doesn’t fit the other categories, such as Lambda functions or internal HTTPS endpoints.

The Approval Workflow

Every record in the Registry follows a strict lifecycle:

• Draft: The initial state, where the record is created but not visible to consumers.
• Submitted for Approval: The record enters the approval pipeline, where designated reviewers assess its compliance with organizational standards.
• Approved: Only records in this state appear in searches and are available for use across the organization.
• Rejected or Deprecated: Records that fail validation or are outdated remain in the system for auditing but are hidden from consumers.

This workflow ensures that every agent, skill, or MCP server meets the organization’s requirements before it’s put into production—eliminating the risk of shadow IT in the AI era.

Technical Requirements and Limitations

Deploying the Registry isn’t without its quirks. Based on hands-on testing, here are the key gotchas to watch for:

• SDK compatibility: The Registry requires boto3 ≥ 1.42.87. Older versions won’t recognize the new methods.
• CLI version: AWS CLI must be v2 ≥ 2.34.28. Running an older version will result in errors like Found invalid choice.
• Preview regions: The service is currently available in five regions: us-east-1, us-west-2, ap-southeast-2, ap-northeast-1, and eu-west-1.
• Manual discovery: Despite its name, the Registry is positioned as a discovery service, not a build tool. Future integrations with Runtime and Gateway are likely but not yet available.

Pricing and Scalability

During the preview phase, the Registry is free to use. When it reaches general availability (GA), AWS plans to charge based on the number of Net Records—active records at any given time. Deleting a record reduces the count, ensuring costs scale linearly with adoption.

Additional services like EventBridge, SNS, and IAM will incur their standard pricing, but for most lab environments, these costs remain negligible—often just cents per month.

A Proactive Approach to AI Governance

The introduction of AWS Agent Registry marks a turning point in the AI adoption lifecycle. For organizations still in the early stages of deploying agents, it offers a chance to implement governance before sprawl becomes unmanageable. For those already grappling with dozens of uncoordinated tools, it provides a path to consolidation and standardization.

The message is clear: agent governance isn’t optional—it’s essential. Whether your organization has 8 agents or 50, the time to act is now. Waiting for the problem to escalate will only make the cleanup effort more costly and disruptive.

As AI continues to evolve, tools like the Agent Registry will become the backbone of corporate innovation—ensuring that the promise of AI isn’t undermined by the chaos of unmanaged growth.