Agentic PAI - AI-Powered Operations Interface - Platform For AI

Key features

Feature	Benefit
Natural language interface	Describe your goal in everyday language—no need to remember API parameters, instance specifications, or console paths. Business users can get started quickly.
Cross-region unified management	Single service entry point for all regions. No need to switch consoles region by region.
Safe and controlled operations	Read and write operations are distinguished. Write operations require explicit confirmation and are governed by RAM permissions.
Native PAI integration	Deeply integrated with PAI capabilities. Operation results sync with the console in real time.
Personalized memory	Remember your default region, frequently used workspaces, and instance type preferences. Subsequent conversations automatically apply these settings.
Scheduled task automation	Configure scheduled tasks for repetitive work like inspections, monitoring, and resource reclamation. The system executes them automatically.
IM notifications	Task results are pushed to DingTalk or Lark groups via webhooks. Critical information is never missed.
Auditable history	Conversation records and operation history are persistently stored for review, troubleshooting, and team handover.

Core capabilities

Conversational task submission

Describe your requirements in natural language. The Agent handles intent recognition, parameter completion, API construction, and task execution.

Execution flow: Intent recognition → Parameter completion (prompts for missing information) → Confirmation (for write operations) → Submit and execute → Display results.

Conversation example

User: Create a DSW instance with an A10 GPU and mount the nas-workspace dataset.
Agent: I need a few details—what should the instance be named? How long should it run?
User: demo-dev, 24 hours.
Agent: DSW instance "demo-dev" (A10 × 1) created. nas-workspace mounted. Starting up—available in about 2 minutes.

Scheduled tasks and IM notifications

Configure scheduled tasks through a form to automatically execute read or analysis operations at specified intervals. View results in task history, or push them to DingTalk or Lark groups via webhooks.

Note

Currently, only read and analysis operations are supported (querying status, computing usage statistics, inspection alerts, etc.). Write operations (create, delete, modify) require confirmation and can't be reliably executed on a schedule, so they aren't supported yet.
After configuration, click Run now to verify the output meets expectations before enabling scheduled execution.

Configure scheduled tasks

Go to the Scheduled Tasks page.
In the upper-right corner, click Create scheduled task.
Fill in the task name, task instructions, trigger frequency (cron expression or UI selection), notification channel, webhook URL, and task description (in natural language).
Click Run now to test and verify.
After verification, enable scheduled execution.

Get a DingTalk webhook: Group Settings → Robots → Add → Custom (Webhook) → Copy the URL.

Intelligent Q&A and diagnostics

Ask questions in natural language. The Agent uses PAI product knowledge and current page context to provide instant responses for problem diagnosis, terminology explanations, and selection recommendations.

Supported scenarios

Scenario	Description
Resource runtime diagnostics	DSW startup failures, SSH connection issues, long training job queues, EAS service deployment failures—automatically identifies root causes and provides fix entry points.
Console error explanations	Provides plain-language explanations and solutions for console popup errors.
Product knowledge Q&A	Real-time terminology explanations, feature limitation inquiries, and operation path guidance.
PPU specialized support	Environment validation, image recommendations, and model recommendations from Model Gallery that are adapted for PPU.
Selection and solution recommendations	Recommends models, instance types, and PD disaggregation configurations based on training or inference scenarios.

Use cases

Platform/resource administrators

Manage PAI cluster resources, workspaces, and infrastructure globally. Focus on resource utilization, cluster health, permissions, and quotas.

Scenario	Example conversation or operation
Resource inspection	Inspect all DSW instances in the Hangzhou region. List abnormal ones and those running for over 24 hours.
Quota monitoring	Show the used/allocated ratio for all quotas in Beijing. Highlight those with usage over 80%.
Node status inspection	Check all Lingjun nodes in the Ulanqab region. List nodes not in Ready state.
Cluster capacity planning	Compute the available GPU count and allocated ratio for each resource group in the Hangzhou region.
Member and permission management	List all members in workspace pai-prod and show their role distribution.
Scheduled inspection alerts	Configure a scheduled task: inspect instance status and node health across all regions daily at 9:00 AM. Send alerts to a DingTalk group.

Algorithm engineers/model trainers

Use PAI daily for model development, training, and deployment. Focus on development efficiency and training stability.

Scenario	Example conversation or operation
Quick development environment setup	Create an A10 single-GPU development machine with PyTorch installed.
Batch training submission	Clone and submit 3 training jobs using dlcxxxxxx, with learning rates lr=0.01/0.001/0.0001 respectively.
Training status query	Where is my DLC job dlc-abc123 in Hangzhou? What's the current loss?
Dataset management	List datasets I can access. Mount nas-shared to my development machine.
Training log viewing	Show the last 100 lines of logs from worker-0 of dlc-abc123.
Training monitoring alerts	Configure a scheduled task: push the loss curve of the training job to a Lark group every 30 minutes.

Model deployment/inference engineers

Focus on EAS model service deployment, operations, and performance optimization.

Scenario	Example conversation or operation
Service deployment	Deploy the xxx model from Model Gallery as an EAS service.
Auto scaling configuration	Adjust the scaling policy for service-rec-001: scale out when GPU utilization exceeds 70%, with a maximum of 5 instances.
Multi-version management	List current versions of service-rec-001. Route 80% of traffic to v2.
Service monitoring	Check the QPS and P99 latency for service-rec-001 over the last hour.
Service health check	List all EAS services and highlight any with abnormal status.
Service decommissioning	Stop v1 of service-rec-001. Keep v2 and v3.

Team managers/cost owners

Focus on resource costs, utilization trends, and team resource allocation from a management perspective. No direct operations—emphasis on data insights.

Scenario	Example conversation or operation
GPU utilization weekly report	Compute the average GPU utilization for all resource groups in the Hangzhou region this week. Rank by resource group.
Team resource consumption	Rank DSW instance runtime for all members in workspace pai-prod.
Resource allocation comparison	Compare resource quotas and actual usage between pai-dev and pai-prod workspaces.
Scheduled cost reports	Configure a scheduled task: push last week's resource utilization and cost report to a DingTalk group every Monday at 9:00 AM.

FAQ

Which PAI modules are supported?

Currently supports DSW, DLC, EAS, workspace, and resource management. More modules are being added.
Which regions are supported?

Centralized service, integrated with all regions in your account. No need to switch region by region.
What operations do scheduled tasks support?

Only read and analysis operations (queries, statistics, inspections, etc.). Write operations require confirmation and aren't supported in scheduled scenarios yet.
What frequencies do scheduled tasks support?

Supports cron expressions. Configure daily, weekly, monthly, or other schedules. The highest frequency is hourly.
Which IM platforms do webhooks support?

Currently supports DingTalk and Lark robots.
Are conversation records saved?

Yes. All conversation history is retrievable for operation review and team handover.
How is this different from general AI like ChatGPT?

Agentic PAI can actually operate PAI resources (create instances, submit training jobs, deploy services)—not just provide documentation or code snippets.
How is security ensured?

Write operations require explicit confirmation. All operations are governed by the executor's permissions on Alibaba Cloud. The Agent can only do what your account is authorized to do.

Pricing and support

Currently in public preview. The PAI team covers token usage costs—you don't pay.

After commercialization, it will be included in the PAI billing system. Pricing is expected to be based on token usage or monthly subscriptions. Final pricing is subject to official announcement.