Published at: 2026-09-17
Agent studio admin guide
This guide is for enterprise admins and business configurators. It explains how to configure, manage, and publish Agents in Agent Studio.
Agent Designer layout
Go to Setup > AI Platform Management > Agent Studio > Agent.
The Agent Designer uses a standard three-column layout with a top bar:
Top bar: multi-version management
| Element | Description |
|---|---|
| Agent name | The Agent being edited. Click to return to the Agent list |
| Version selector | Shows the version being edited (such as V1.0). Switch versions from the drop-down. The active version is the one taking effect for users |
| Use This Version | Shown for non-active versions. Click to switch this version to active |
| Copy as New | Copy the current version to a new version. Use this to debug a new version, then switch it to active and publish it |
| Delete | Delete a non-active Agent version |
| Exit | Exit the Agent Designer |
Sidebar overview
| Menu | Purpose |
|---|---|
| Basic Info | Configure Agent basic info, Agent settings (system prompt), model, and welcome message |
| Skills | Select Skills from the enterprise Skill library to mount so the Agent can discover and use them |
| Knowledge Base | Bind an enterprise knowledge base so the Agent can search internal documents when answering |
| Context | Manage variables injected into the Agent context (system, channel, and custom variables) |
| Event Log | View run records and tool call details for each Query |
Debug preview
| Feature | Description |
|---|---|
| Real-time conversation | Enter a question to view the draft Agent’s reply, Agent Loop reasoning, and tool call details |
| Version switching | After you switch versions on the top bar, the debug area reflects the change immediately |
| Context Mock | Mock context content for conversation testing |
Agent basic info
Page sections
| Section | Fields | Description |
|---|---|---|
| Details | Name, API Name, Agent settings, description, applicable scope, type, avatar | Agent identity |
| Model info | Model provider, model selection | Agent brain |
| Welcome message | Welcome title, welcome message, quick questions | Onboarding message shown when users first enter |
Field details
Details
| Field | Required | Description |
|---|---|---|
| Name | Yes | Display name, visible to users |
| API Name | Yes | Unique identifier used in system calls |
| Agent settings | Yes | The system prompt that defines the Agent’s persona, boundaries, and style |
| Description | No | Internal note to help admins understand the Agent’s purpose |
| Type | Yes | The Agent type, which determines its use case |
| Applicable scope | Yes | Enterprise employees who can use the Agent |
| Avatar | No | Visual marker to distinguish Agents |
Model info
| Field | Description | Common values |
|---|---|---|
| Model provider | The vendor behind the LLM | Volcano Engine / Zhipu / Alibaba / DeepSeek |
| Model selection | The specific model version | deepseek-v4-pro, kimi-k2.6 |
Welcome message
| Field | Description |
|---|---|
| Welcome title | Eye-catching text shown when users first enter the Agent |
| Welcome message | Body text that guides users to start a conversation |
| Quick questions | Preset recommended questions covering the most common business scenarios |
Tip: A welcome message should state what the Agent can do + one example question to lower the first-use barrier.
Agent settings (system prompt)
Agent settings act as the Agent’s constitution. They define behavioral boundaries and execution style.
| Module | Content | Why it matters |
|---|---|---|
| Identity | What role the Agent plays | Determines its name, personality, and tone |
| Capability | What the Agent does and which problems it solves | Determines its expertise |
| Responsibility boundaries | What it can and cannot do | Prevents overreach |
| Tool preferences | Which Skills or tools to prioritize | Unifies execution style |
| Safety rules | Hard limits that must not be crossed | Last-line protection |
Skills
What is a Skill
A Skill is a packaged business capability that an Agent can call. It captures how to do a task as a resource that a model can recognize and invoke. Skills are modular building blocks of enterprise business know-how.
A Skill exists as a directory in the file system:
skillname/
├── SKILL.md ← Core entry (AI must read)
├── References/ ← On-demand documents
├── Assets/ ← Templates, fonts, and other output resources
└── Scripts/ ← Executable code for deterministic tasks
Problems that Skills solve:
| Problem | How Skills solve it |
|---|---|
| Knowledge is scattered and cannot be reused | Encapsulate tool usage patterns into reusable Skill modules; define once, load on demand |
| Multi-step tasks are error-prone | Skills embed steps and preconditions, reducing hallucination |
| Capabilities are hard to compose and extend | Skills work like Lego bricks; new tasks assemble them without affecting other Skills |
| Context bloats and token costs rise | Progressive disclosure loads only the Skill for the current step |
Add Skills to an Agent
Click Add from Skill Library and select Skills from the enterprise library to mount on the Agent.
Create a Skill
Manage Skills in the Agent Skill Library at Setup > AI Platform Management > Agent Studio > Agent Skill Library.
You can create Skills in two ways: AI creation and import.
Create a Skill with AI
Click Create Skill with AI to open an AI chat box. Describe the business function in natural language. The AI combines the business scenario with built-in tools and ShareCRM CLI methods to create the Skill and upload it to the enterprise library.
Import a Skill
Import a Skill zip package. This is useful for migrating Skills downloaded from external platforms into ShareCRM.
Note: TheSKILL.mdfile in the uploaded Skill must includenameanddescriptionin its YAML frontmatter. Refer to the Claude Skill specification.
Edit a Skill
On the Skill edit page, you can edit Markdown files inline. You can also import (import update) and export a Skill. Each save creates a new version for traceability and rollback.
You can also reopen the Create Skill with AI chat box in Setup and use natural language to ask the AI to modify the Skill.
Agent tools
ShareAgent interacts with the outside world through tools. When a user makes a request, the Agent:
- Decides what task to complete.
- Selects the best tool from available tools.
- Fills parameters based on user input.
- Calls the tool and observes the result.
- Loops until the task is complete if multiple steps are needed.
Tools fall into two categories:
- Built-in tools: Platform-provided capabilities such as file operations, search, rendering, and memory.
- Operation library tools (sharecrm CLI): Tools that connect to ShareCRM business objects, tenant custom APIs, and third-party APIs.
Built-in tool list
| Category | Tool | Purpose |
|---|---|---|
| CRM business | sharecrm |
Unified business CLI to query, create, and update CRM data (accounts, opportunities, contracts) |
| Skill management | read_skill, prepare_skill_improvement, copy_as_personal_draft, save_skill |
Load, modify, and save Skill definitions |
| Scheduled tasks | schedule-def-create, schedule-def-update, schedule-def-enable, schedule-def-delete, schedule-def-list |
Full lifecycle management of scheduled tasks |
| Pipeline | stage-move-to |
Advance pipeline stages |
| File operations | read_file, write_file, edit_file, list_files, file_download, file_upload, file_parse |
Read, write, upload, download, and parse files |
| Content rendering | render_content |
Render results to the front end in real time |
| Memory | memory, memory_query, memory_create, memory_update, memory_delete |
Long-term memory management |
| Shell | shell |
Run shell commands |
| Interaction | AskUserQuestion |
Ask users to collect information |
| Task management | write_tasks |
Multi-step task tracking |
Agent operation library
sharecrm CLI
All
sharecrm commands are managed in the Agent Operation Library at Setup > AI Platform Management > Agent Studio > Agent Operation Library.
The Operation Library maps scattered enterprise capabilities to a unified
sharecrm CLI. Agents discover, call, and complete parameters through one protocol, so enterprises do not need to build adapters for each business action.
Create or edit an operation
In the Operation Library, click New, select the operation type, and fill in the configuration.
The Operation Settings panel has three sections: Operation Info, Category & CLI, and Input/Output Parameters.
Operation Info
| Field | Description |
|---|---|
| Operation type | The type, which determines which metadata the operation can bind |
| Select metadata | The metadata bound to the operation (function, flow, prompt, knowledge base scope) |
| Name | Display name of the operation |
| API Name | Unique identifier of the operation |
| Instruction | Semantic note that explains what the operation does and when to run it |
| Show execution status text | Whether to show status feedback during execution (such as “Searching knowledge base…”) |
| Execution status text | The status text content, used together with the previous option |
Category & CLI
| Field | Description |
|---|---|
| Level 1 category | The business category, such as Knowledge & Retrieval or CRM & Sales |
| Level 2 category | The subcategory, such as Knowledge or Object Operations |
| Method | The CLI method name (such as helpcenterRetrieval). This is the interface identifier ShareAgent calls |
CLI naming convention: Clear, business-meaningful category and Method names help the Agent find and call the right CLI method at runtime.
Input/Output Parameters
| Field | Description |
|---|---|
| Field selection | The parameter name. The UI auto-populates parameters based on the bound metadata |
| Data type | The parameter data type, auto-populated |
| Required | When checked, the Agent must provide a value for this parameter |
| Instruction | Natural-language note that helps the Agent understand what to pass |
Knowledge base management
An Agent’s LLM cannot know your product manuals, price books, or contract templates. A knowledge base is the standard way to feed enterprise knowledge to an Agent.
| Source | Use case |
|---|---|
| Existing enterprise knowledge base | Select an enterprise knowledge base space |
| Local file upload | Product manuals, talk tracks, PDF contracts, FAQs |
Context management
Context is the environment information an Agent senses during a task. Context variables are injected into the Agent system prompt as key-value pairs.
| Category | Description | Typical fields |
|---|---|---|
| System variables | Runtime information injected by the system | Current time, current date, SessionId, chat history, chat messages, enterprise account |
| Channel variables | Channel information from different entries | Chat group info, current page object data |
| Custom variables | Business variables configured by admins | Custom employee info, custom query conditions |
Event log
The event log is the core entry point for troubleshooting, tuning, and audit. ShareAgent records each user Query as a complete event, including:
- User input
- Model reasoning
- Tool call chain (input parameters and output results)
- Final answer
- Session start time
- Duration and token consumption
You must enable the event log in Agent Basic Info before records are kept.
Agent discovery
All internal enterprise Agents can be discovered and called as SubAgents by ShareAgent (the main Agent). The Description field summarizes what an Agent does in one or two sentences. This is the main Agent’s first basis for deciding whether to call it.
