The scale of AI applications is growing rapidly, and so are the management needs around Prompts and Skills. The community already has plenty of Prompt platforms and Skill marketplaces that solve sharing and discovery. But when AI applications enter enterprise production environments, the requirements change—who manages versions? Who approves releases? Can incidents in production be rolled back in seconds? How do different teams isolate their assets? How is access controlled?
These are questions community platforms do not need to answer, but enterprises must.
Enterprise AI assets—Prompts, Skills, and private API integration logic—are scattered everywhere, versions are messy, and unified management is lacking. Code has Git, packages have npm, containers have image registries, and enterprise AI assets likewise need a self-controlled registry.
Today, AI Registry officially opens its public beta to provide enterprises with that answer.
AI Registry is a fully managed AI asset registry launched by Alibaba Cloud, and the cloud SaaS version of Nacos AI Registry capabilities. Built on Nacos under the hood, clients can connect directly using the Nacos SDK, so teams already using Nacos can get started with zero learning cost. It provides unified registration, version management, discovery, and governance capabilities for AI assets such as Prompts, Skills, and Agents, helping enterprises establish a standardized AI asset management system.
Fully managed, zero operations. No need to purchase instances or set up clusters. Create a workspace to get started, and underlying resources are automatically provisioned and scaled. Developers can focus on the AI application itself without worrying about infrastructure operations.
Console management, client integration. Operations such as creating, editing, releasing versions, and managing tags for AI assets are done through the console, which provides a visual management interface. At runtime, applications pull resources in real time through a client SDK or API, separating management from consumption so each can do its own job.
Workspace isolation. Each team or project has an independent workspace with strict resource isolation. Development and production environments are independent and do not interfere with each other. This meets the management needs of multiple teams, multiple projects, and multiple environments.
Enterprise-grade security. Workspace data isolation and fine-grained sub-account permissions ensure strict isolation between different users and different workspaces. Fine-grained access control is supported down to a specific type of resource under a workspace, meeting the permission management needs of different roles within the enterprise.
AI-assisted capabilities. Provides intelligent Prompt optimization and online debugging capabilities to lower the barrier to Prompt engineering.
Prompt is the most core and most frequently changing asset in AI applications. The behavior of an Agent depends largely on the quality of the Prompt. But in real development, Prompt management is often the most rudimentary—written as code constants, requiring a full release cycle for every change; or stored in configuration files, where version conflicts keep arising during collaboration. Even more troublesome is production operations: which Prompt version is running in production? Who made the last change? Can a bad new version be rolled back quickly? Without dedicated tools, these issues can only be handled through manual records and verbal agreements.
The first batch of Prompt management capabilities in the AI Registry public beta is designed to upgrade this process from manual mode to engineering-grade management, covering the full chain from authoring and debugging to production runtime.
Each Prompt has a complete version history, with draft, review, and release states clearly distinguished. Releases use a tag-routing mechanism, and tags are defined by users themselves; common tags include stable, canary, and gray. During canary validation, point the canary tag to the new version while production traffic continues to use stable. After validation passes, switch over. If a problem occurs, switch the tag back to the previous version for a rollback in seconds, with no impact on the business.
At runtime, applications pull Prompts by tag or version number through SDK or API, with support for incremental checks—if there is no change, it will not fetch again. After a Prompt is updated and released on the platform, clients become aware of it in real time without needing to redeploy the application.
In the past, changing a Prompt required going through the code release process. Now, editing and releasing are done on the platform, taking effect instantly in production and greatly shortening the Prompt iteration cycle.
The platform has built-in multi-model streaming debugging capabilities. After writing a Prompt, you can directly choose a model in the console, adjust parameters, and verify results, which is ideal for quick validation during daily editing. It also provides AI-assisted optimization that automatically analyzes the Prompt and offers improvement suggestions, lowering the barrier to Prompt engineering. If you need more systematic multi-version comparison experiments and automated evaluation, you can use AgentLoop's Playground (see the next section).
Going forward, Prompt management will also integrate security review capabilities, automatically detecting prompt injection risks, sensitive information, and other security issues before release, adding another line of defense for enterprise production environments.
AgentLoop is Alibaba Cloud's AI Agent full-lifecycle data observability and performance optimization platform. Centered on Trace, Log, Metric, Conversation, and other data generated at AI application runtime, it provides end-to-end observability, automated evaluation, Playground experiments, and long-term memory capabilities, helping enterprises build a continuous improvement loop for Agents.
Prompt optimization is not a one-time task. Writing the first version and deploying it is only the starting point; what really determines the effectiveness of an AI application is continuous iteration afterward—observing production performance, finding problems, making adjustments, and redeploying. The faster this loop runs, the better the application performs.
After AI Registry and AgentLoop are connected at the Prompt management layer, they form a complete data flywheel chain.
AgentLoop's Playground is a visual experimental environment that supports configuring up to 5 experimental branches at the same time, comparing the performance of different models, different parameters, and different Prompt versions.
Playground can directly select Prompts managed in AI Registry without manually copying content, and version alignment is completed automatically. Use the {{variableName}} syntax to inject dataset samples, mount evaluators for automatic scoring, and quantify the performance differences of each version. The system records a complete snapshot of each experiment—model configuration, token consumption, first-token latency, and evaluation scores—ensuring the experiment results are traceable and reproducible.
After confirming the optimal version, release it in AI Registry and control the canary ratio through tag routing (stable / canary). Online applications pull the latest version in real time through the SDK without redeployment.
After a Prompt goes live, AgentLoop automatically collects production Trace data through the Python probe—every LLM call's input and output, the Agent execution chain, tool invocation details, token consumption, and response latency are all recorded in the call chain.
Based on this production data, AgentLoop's evaluation tasks can run continuously: automatically evaluating production Traces by sampling rate, scoring them in real time with preset evaluators, covering general scenarios (safety, coherence, completeness), RAG scenarios, and Agent scenarios, and automatically identifying low-scoring samples and Bad Cases.
Bad Cases discovered in production do not stay in reports. Import the low-scoring samples into a dataset, adjust the Prompt in the Playground in a targeted way, set a baseline group through the comparison analysis feature, calculate the difference values of the new version across dimensions such as accuracy, latency, and cost, and use text-level Diff highlighting to locate output changes. After confirming the improvement, release the new version to AI Registry, and the online application will pull it automatically.
The entire chain forms a closed loop: Create → Debug and evaluate → Canary release → Production observability → Problem discovery → Iterative optimization. AI Registry is responsible for version management and release control— which version is in canary, which is in production, and switching in seconds when a rollback is needed. AgentLoop is responsible for performance measurement and problem discovery—how it is running in production, where it is degrading, and whether things have actually improved after changes. Data keeps flowing through the chain, and Prompt quality improves with each cycle.
Going forward, AgentLoop's observability capabilities will be further integrated with AI Registry—by collecting Prompt names and version information from the client, developers will be able to directly see the traffic distribution and performance metrics of each online version, further shortening the path from "discovering a problem" to "pinpointing the specific Prompt version."
In addition to Prompt management, AI Registry's Skill management capabilities will officially launch at the end of April.
A Skill is the smallest reusable unit of Agent capability—a tool for handling date formats, a wrapper for calling internal APIs, or a piece of domain-specific reasoning logic can all be packaged as a Skill for multiple Agents to discover and reference.
The current AI Skill ecosystem is growing rapidly, but enterprise scenarios face several real-world problems:
AI Registry's Skill management is designed for enterprise scenarios and provides complete capabilities from registration to runtime:
Registration and format compatibility. Skills are organized as ZIP packages containing a SKILL.md description file and resource files (templates, data, scripts, etc.). This format is compatible with community Skills, so existing Skills can be imported directly, and they can also be created and edited online through the console.
Version management and canary releases. Like Prompt, Skills support complete version history and custom tag routing. New versions are first tagged canary for validation in a canary environment, then switched to stable after passing; if problems occur, rollback can happen in seconds. The change history for each version is fully retained to meet audit and traceability requirements.
Review workflow. Skill releases are no longer "submit and go live." A draft → review → release workflow is supported, and Skills for production environments must go through explicit approval. Test environments can be configured for direct release, while production environments require mandatory review, balancing development efficiency and release safety.
Independent activation and deactivation control. Supports independent activation and deactivation at both the Skill level and the version level. You can take an entire Skill offline, or only take a problematic version offline, without affecting the normal operation of other versions.
Search, discovery, and client consumption. Skills registered in the Registry support fuzzy search and tag filtering, making it easy for teams to discover and reuse them internally. The client search interface only returns Skills that are enabled and have an online version, allowing Agents to dynamically discover and load the capabilities they need at runtime.
Download statistics. The download count of each Skill version is automatically tracked, helping teams understand which Skills are widely used and which need promotion or retirement.
Discover useful Skills from the community, bring them into your own enterprise registry, and decouple from external platforms—this is the Skill governance path AI Registry provides for enterprises.
AI Registry officially opened its public beta on April 15. Log in to the console and create a workspace to start managing your AI assets.
AI Registry Console: https://mse.console.alibabacloud.com/#/ai-registry
AI Registry Product Documentation: https://www.alibabacloud.com/help/en/mse/user-guide/ai-registry-ram-permission-configuration-guide
AgentLoop Documentation: https://www.alibabacloud.com/help/en/cms/cloudmonitor-2-0/what-is-agentloop
695 posts | 56 followers
FollowAliware - February 5, 2021
Alibaba Cloud Native Community - March 13, 2026
Alibaba Cloud Native Community - September 19, 2023
Alibaba Cloud Native Community - February 3, 2026
Alibaba Cloud Community - July 11, 2024
Alibaba Developer - March 8, 2021
695 posts | 56 followers
Follow
Container Compute Service (ACS)
A cloud computing service that provides container compute resources that comply with the container specifications of Kubernetes
Learn More
Container Service for Kubernetes
Alibaba Cloud Container Service for Kubernetes is a fully managed cloud container management service that supports native Kubernetes and integrates with other Alibaba Cloud products.
Learn More
Microservices Engine (MSE)
MSE provides a fully managed registration and configuration center, and gateway and microservices governance capabilities.
Learn More
Tongyi Qianwen (Qwen)
Top-performance foundation models from Alibaba Cloud
Learn MoreMore Posts by Alibaba Cloud Native Community
