Agent

Prompt Agent vs orchestrated Agent · Tools · Knowledge base · Debugging · Publishing

An Agent is Evose's conversational AI entity. Two paths share one capability set.

Two Kinds of Agent

Type	How to build	Best for
Prompt Agent	A role prompt + pick a model + pick tools + pick knowledge bases	Simple customer service, assistants, translation, writing
Orchestrated Agent (Chatflow)	Visual node canvas + 9 groups, 19 nodes	Multi-turn forms, complex CS, guided dialog

Workspace → Apps · Agent → New → Prompt Agent
Fill in metadata: name, icon, description
Edit the role prompt:
You are the {role} for {company}. Your task: {task description} Rules: - {rule 1} - {rule 2}
Choose a base model (inherit org default or specify explicitly)
Configure sampling: temperature / TopP / max length
Bind a knowledge base (optional, multi-select)
Bind tools (optional, multi-select)
Configure memory:
- Session memory: context within the current conversation
- Long-term memory: cross-session user preferences (optional, must be enabled)
Save → debug → publish

Group	Node	Purpose
AI	LLM	Call a language model
AI	Knowledge retrieval	Retrieve top-K from a knowledge base
Logic	Condition	if/else branching
Logic	Batch	Process a list item by item
Logic	Loop	Continue while condition holds
Logic	Intent classification	Classify user input into a category
Data	Code	Run a Python/JS snippet
Data	Template	String template formatting
Data	Variable	Set/get variables
Data	Document extraction	Extract text from a file
Data	Parameter extraction	Extract structured params from dialog
Network	HTTP	Call any REST API
App	Agent	Call another Agent
App	Workflow	Trigger a Workflow
Tool	MCP / Plugin / Marketplace	Call external tools
Interaction	Direct reply	Send a fixed message to the user
Interaction	Form input	Pop a form to collect structured input
End	End	Terminate the round

In conversation, an Agent can automatically decide which tool to call (LLM function calling).

Tool source	How to integrate
MCP	Standardized protocol; recommended. Provide a URL or launch command
HTTP plugin	RESTful API + auth + schema configuration
EvoTool marketplace	Out-of-the-box tools maintained by Evose and the community

When you bind one or more knowledge bases, the Agent retrieves before answering each turn:

Behavior	Description
Passive retrieval (default)	Each user message triggers retrieval
Active query	Prompt explicitly says "use the knowledge retrieval node to look up X first"
TopK	Default 5; adjustable in Agent config or Chatflow node

The Debug panel on the right of the Agent edit page:

Chat directly to validate response quality
See each message's Trace: retrieval results, tool calls, LLM in/out, token counts
Switch models for comparison
Edit prompts and hot-reload without saving to test instantly

Once published to the Workbench, authorized users find it in Workbench · Task.

Setting	Description
Visibility	Just me / specific members / roles / departments / entire workspace
Channels	Workbench · API · Web embed · Mobile SDK · DingTalk · WeCom · Feishu · WeChat · Mini-program
Versioning	Each publish creates a new version; rollback supported

Channel	Integration
Web	JS SDK / iframe
Mobile	iOS / Android SDK
Enterprise IM	DingTalk · WeCom · Feishu · WeChat · WeChat Official Account · Mini-program
API	RESTful, see API docs

After an Agent goes live, see in Workspace Observability · Resources: