This document describes how to publish and deploy Agents in DataBuddy, covering deployment sources, deployment processes, runtime status monitoring, and version management and download.
Overview
Agents are deployed and run on the platform as Apps, with code, dependencies, runtime status, versions, and permissions managed centrally by the App. Deployment from Studio is currently supported, and deployment from Git will be supported in a later version.
After deployment, the Agent can be invoked through ChatUI, APIs, or by other Agents.
Prerequisites
The code for the high-code Agent has been developed and basically debugged in Studio.
The associated MLflow experiment has CAN_EDIT permission.
The platform base model is already on the allowlist.
Use Limits
Limit | Description |
Deployment Source | Currently, deployment is supported only from Studio. Deployment from Git will be supported in a later version. |
Deployment File Management | Supports downloading the Agent code package by clicking. |
Platform Foundation Model | Enabled through allowlist, with usage tracking supported. |
Billing | Billing is not available currently. A separate notice will be provided after the official commercial launch. |
Operation Steps
Step 1: Develop and Adjust Code in Studio
1. Go to the Agent overview page and click the Code Development module shortcut, or click the Code Development Tab in the top navigation.
2. Studio automatically opens the code template associated with the Agent (OpenAI / LangGraph).
3. In Studio:
Edit and debug the code.
Modify the environment variables.
Use Studio's AI-assisted capabilities, such as code generation and tool discovery, to facilitate development.
Note:
Code synchronization rules:
The code in Studio is synchronized to COS.
When you click the download button on the Agent details page, the platform calls the Studio API to pull the latest code and package it for download.
Deployment always uses the latest code in Studio as the source of truth. If users make local modifications and upload them, the changes should be synchronized to Studio.
Step 2: Trigger Deployment
In the status card at the top of the Agent details page, click the Deploy button. This button is clickable only when the status allows deployment. For more information, see the Status Transition section. The platform performs the following actions:
1. Pull the latest Agent code from Studio.
2. Install code dependencies and apply environment variables.
3. Start the Agent runtime instance.
4. The call Endpoint is automatically registered and can be invoked by ChatUI, APIs, and other clients.
The deployment progress is displayed in real time through a progress bar.
Note:
During deployment, the Agent cannot provide external services. To ensure availability, schedule deployments during off-peak hours or notify users in advance.
Step 3: Observe Running Status on the Overview Page
After deployment is complete, return to the Agent Overview page to view:
Metric | Description |
Calls Today | Total number of requests from 00:00 to the current time (including successful / failed requests). Click to go to Tracing and Evaluation. |
Success rate. | Successful responses / Total requests × 100% (success = HTTP 2xx and normal response). Click to go to Runtime Monitoring. |
Average latency. | Arithmetic mean of end-to-end response time for successful requests. Click to go to Runtime Monitoring. |
Current Version | Deployment version number running online. Click to go to Deployment Status. |
Token Consumption Today | Total Token consumption of all traces on the current day, displayed in thousands (k / m / b). |
Token Consumption in Last 30 Days | Total Token consumption of all traces in the last 30 days, displayed in thousands. |
Step 4: Test the Deployed Agent
Test by using any of the following methods:
Method | Description |
ChatUI preview experience | Chat directly on the right side of the Agent overview page |
API call | Initiate an HTTP request through the Endpoint URL (view/copy it in the info pop-up on the Agent overview page). |
Step 5: Download the Code Package if Necessary
When you click Download on the Agent Overview page or in the Code Development module, the platform pulls the latest code from Studio, packages it into a zip file, and downloads it to your local machine.
Status Transitions
An Agent has the following states during its lifecycle:
Current status | Available Operations | Next state |
Stopped | Starting | Starting |
Starting | - | Deploying (only after the agent is created for the first time) / Running |
Deploying | Deployment | Running (succeeded) / Previous state (failed) |
Running | Stop / Deploy | Stopping / Deploying |
Stopping. | - | Stopped. |
Error | Start / Stop | Starting / Stopping |
The page displays the corresponding run view and available actions for each state. When the Agent is in the Error state, clicking the error message shows the specific error cause.
Deployment Progress Display
During deployment, the page displays a progress bar with stage-by-stage labels:
Pull Code
Installation Dependency
Start Instance
Register Endpoint
If any stage fails, the system rolls back to the previous state before deployment and retains error information for troubleshooting.
Permission Management
Configure deployment-related permissions in the upper-right corner of the Agent details page:
Permission Point | Description |
Available | Can call through chat and view traces |
Manageable | Can trigger deployment/start/stop, download code, and modify configurations |
The authorization scope covers members across workspaces under the entire root account.
Version Management
After the Agent is deployed, each redeployment generates a new version:
The Current Version section on the overview page displays the version number running in the production environment.
Click to go to the deployment status page to view the historical version list.
Note:
The current version management is mainly used to trace which deployment is currently running in the production environment. Full version rollback, multi-version canary release, and other capabilities will be available based on the release schedule of subsequent versions.
Running Monitoring
Go to the Deployment Ops Tab on the Agent details page to view:
Module | Description |
Running Monitoring | Call volume trends, success rate trends, and average latency trends |
Deployment Status | Current deployment progress and historical deployment records |
Call Methods
After successful deployment, the Agent provides the following invocation methods:
Method | Scenario | Description |
ChatUI | Direct conversations for business users | A frontend template preset by the platform that developers can further develop. |
HTTP API | Integrate into your own application | Call through Endpoint URL + authentication |
Model Service Node | Mutual calls between agents / calls in SQL Editor | Call through the MCP protocol and configure accordingly. |
FAQs
Q: How Do I Troubleshoot Deployment Failures?
A: Common causes include: ① Code dependency installation failure (check the error message on the deployment progress bar). ② Insufficient MLflow experiment permissions (confirm that the account has CAN_EDIT permission on the experiment). ③ The platform base model allowlist is not enabled (contact the administrator). ④ The called MCP service is unavailable (check the service status in the MCP Overview list). Q: How Do I Roll Back to the Previous Version?
A: Currently, redeployment is mainly supported. If a rollback is required, the following is recommended: ① Use Git in Studio to switch to the code of the corresponding version. ② Alternatively, download the code package of the previous version, modify it locally, upload it back to Studio, and then redeploy. Full version management and rollback capabilities will be available based on the release schedule of subsequent versions.
Q: Is Canary Deployment Supported?
A: Currently, canary release is not supported. It is recommended to use a separate "Test Agent" during the testing phase, and deploy it to the production Agent after verification is complete. Multi-version and canary release capabilities will be introduced in subsequent versions.
Q: What Should I Do If the Agent Is in the "Running" State but Not Responding to Calls?
A: First, check on the overview page whether "Today's Calls" and "Success Rate" have dropped abnormally. Then, check the recent Trace status in the Trace list to identify the specific failure cause (mostly tool call failures or model unavailability). If the status metrics are abnormal, stop the Agent and then restart it.
Q: When Will Deployment from Git Be Supported?
A: Currently, only Studio deployment is supported. Git deployment will be provided in subsequent versions, with the specific timing subject to the platform Roadmap announcement.
References