Basic information configuration

Click the Edit icon on the right side of My Agent in the upper left corner of the page to edit the agent's name, description, and avatar.

• Application name: Required, can be customized naming, character limit of 50.
• Application description: Required, the ability description of the agent, it is recommended to describe the ability range of the agent as clearly as possible. In the future, if other autonomous planning agents or intelligent intent planning engines schedule the agent, the description will be used as the capability positioning for planning selection, with a character limit of 500.
• Avatar: Generated by the system by default, you can also click to randomly select to switch the new avatar.

Model configuration

• Support turning on multi-modal recognition capability, after turning on, it will support multi-modal input of pictures.
• Planning Model: Select the specified model for autonomously planning execution steps based on user input intentions.
• Planning prompt word: the agent's planning prompt word, it is recommended to fill in the agent's role, task, execution requirements, restrictions and other rules.

Tool

Support adding MCP tools (tools under all MCP Servers registered in MCP), agent tools (external agents or self-built agents in the agent center), SKills (published Skills in the Skill Center).

Add an MCP tool or agent tool:

1. Click + to add an MCP tool or an agent tool. See the creation of self-built MCP tools MCP.
   
2. Click the Edit button with the added MCP or agent card to edit the capability description of the MCP tool or agent under the current agent. The modified capability description is only used by the agent and will not affect the description configuration of the tools and agents.
   
3. Click the Parameter Configuration button where the MCP or agent card has been added to configure the [Pre-Execution Function], [Post-Execution Function] and [Tip Word Enhancement Rules] executed by the tool.
   
   • Pre-execution rules: Some rule actions performed before the tool is executed, often used for data conversion or processing.
     For example, in the stock analysis scenario, the user's query is < XX market >, and the tool imported parameter is required to be the stock code, then you can specify in the pre-function what service should be called for stock code acquisition
   • Post-execution rules: Some rule actions performed after the tool executes the information acquisition are often used to do information format processing, which is convenient for large models to better understand the information obtained.
     For example, in a financial advisory scenario, the content returned by the data query tool may appear in English or the format is not suitable for the language model to understand. At this time, if the tool is directly exported parameters for the next step, it is very likely to cause misunderstanding. Therefore, the data can be formatted in the tool's post-function. Resp.content = "Get the market. The tool returns:" + Resp.content #The tool returns the content field may be in English or the format is not suitable for the model to understand. It needs to be formatted in the post-function
   • Tip Enhancement Rules: A large model attention enhancement solution proposed by Claude Code, when dealing with complex tasks, the large model will forget instructions or plans, through prompt word enhancement, reminding the model of the task that should be done after the tool is executed. It can be configured either by prompt or by code.
     For example, in the stock analysis scenario, the user's query is < XX market >, and the tool imported parameter is required to be the stock code, then you can specify in the pre-function what service should be called for stock code acquisition

Add Skills:

1. Click + to add the required Skills. See the self-built Skill Management operation Skill.
   
2. The added skills will be displayed in the Skills area, which supports deletion operations.
   

Knowledge

Add knowledge. Click the + icon to add three types of knowledge: term base, QA question and answer base, and general knowledge base, which can be used as additional information input when the agent executes to help the agent better generate answers. Multiple knowledge bases of each type can be added. The creation of the knowledge base can refer to the enterprise knowledge engine Create a New Knowledge Base.

Memory

Memory allows users to configure their conversation rounds, short-term memory, and long-term memory on demand.

• Short-term memory: On means that the original text of the historical session and the overview of the historical session are given to the agent as context input.
  • Number of dialogue rounds: The number of context rounds recalled when intent recognition is rolled, and the context dialogue information is used as supplementary information during query rewriting and intent recognition.
• Long-term memory: Automatically retrieve records of user portraits, preferences, and events.
  • Memory: Specifies the memory currently required from the session memory.
  • Memory usage rules: support multiple selection
    • Read Memory: Allows the agent to read historical memories in this memory bank.
    • Write Global Shared Memory: Allows the agent to precipitate the content of this conversation into a global generic portrait memory/event memory.
    • Write to agent memory: Allows the agent to precipitate the content of this conversation into a memory only for the current agent. After selecting this rule, you need to specify a memory retrieval policy.

Dialog configuration

1. Opening remarks: the fixed guide content of the opening remarks;
2. Opening line problem: Configure the fixed problem of opening line guidance, add up to 5;
3. Dialogue guide question: closed by default, the dialogue guide question is the related question after the answer output. Output rules can be configured by prompts;
4. Prompt: Fixed prompt that can be configured to appear before or after the answer output. It is usually displayed as a risk warning after the answer output.
5. Show the source of the question: After opening, the content obtained through the knowledge base or online search tool in the answer will display the file or website address of the corresponding content through the corner mark.

Lifecycle Dynamic Function

An open code node, which can input the rule requirements of the agent before execution, before thinking, after thinking, before planning, during planning, and after planning

Streaming output filter function

The middleware code framework for streaming output, the framework guarantees that each chunk will pass through two of the hooks, where you can decide it "release/replace/discard", and can do cumulative state processing across chunks by custom_context .

Security policy configuration

Support for large model application firewall functionality with BytePlus to ensure model input and output content is safe, available, and trusted.

• How to obtain parameter configuration information such as AK, SK, server address, firewall ID, region, etc., please refer to the description of key parameters of large model application firewall.
• Whether to verify the request content: After opening, the request content will be verified for security compliance.
• Whether to verify LLM output: After opening, the large model output will be verified for security compliance.
• Hit risk control prompt:
  • Content generation using security services: Prompts are generated using the internal capabilities of a large model application firewall.
  • Use a custom security prompt: You need a custom prompt in the input box.

Timeout and limited viewership of

• Speed Limit Configuration: The maximum number of requests per second (or minute), the default is null to indicate no limit.
• Error code: The error code returned when the limit viewership is reached, the default is 429.
• Error message: The prompt message returned when the limit viewership is reached, the default is "There are too many visitors at present, please try again later".
• Sub-agent timeout time: Set the timeout time of calling the downstream sub-agent, in seconds.

Basic information configuration

Click the Edit icon on the right side of My Agent in the upper left corner of the page to edit the agent's name, description, and avatar.

• Application name: Required, can be customized naming, character limit of 50.
• Application description: Required, the ability description of the agent, it is recommended to describe the ability range of the agent as clearly as possible. In the future, if other autonomous planning agents or intelligent intent planning engines schedule the agent, the description will be used as the capability positioning for planning selection, with a character limit of 500.
• Avatar: Generated by the system by default, you can also click to randomly select to switch the new avatar.

Model configuration

• Support turning on multi-modal recognition capability, after turning on, it will support multi-modal input of pictures.
• Thinking Model: Select the specified model for logical derivation and output the thinking logic for solving the problem. Expand this configuration item to support enabling:
  • Deep thinking: After opening, the model will use depth thinking mode.
  • Model generation thinking process: After opening, you need to configure the thinking process prompt word, and use the thinking process prompt word to let the model generate a thinking process text, which will enter the internal arrangement logic of the agent and affect the final answer.
• Planning model: Select the specified model and configure the prompt word, which is used to plan the execution steps independently according to the intention of the user input.
• Reflection Model: Select the specified model and configure the prompts for logical derivation to output the reflection logic for problem solving.
• Reporting model: Select the specified model and configure the prompt word to output the reporting logic to solve the problem, and support streaming output.
• Model cover bottom:
  • Thinking Bottom Model: Alternative when the main thinking model is not available
  • Deep thinking: After opening, the model will use depth thinking mode.
  • Planning, Reflection, Reporting Bottom Model: A Universal Bottom Model for Stages Other than Thinking

Tool

Support adding MCP tools (tools under all MCP Servers registered in MCP), agent tools (external agents or self-built agents in the agent center), SKills (published Skills in the Skill Center).

Add an MCP tool or agent tool:

1. Click +, you can choose to add MCP tools or agent tools, about the creation of self-built MCP tools can be seen MCP.
   
2. Click the Edit button with the added MCP or agent card to edit the capability description of the MCP tool or agent under the current agent. The modified capability description is only used by the agent and will not affect the description configuration of the tools and agents.
   
3. Click the Parameter Configuration button where the MCP or agent card has been added to configure the [Pre-Execution Function], [Post-Execution Function] and [Tip Word Enhancement Rules] executed by the tool.
   
   • Pre-execution rules: Some rule actions performed before the tool is executed, often used for data conversion or processing.
     For example, in the stock analysis scenario, the user's query is < XX market >, and the tool imported parameter is required to be the stock code, then you can specify in the pre-function what service should be called for stock code acquisition
   • Post-execution rules: Some rule actions performed after the tool executes the information acquisition are often used to do information format processing, which is convenient for large models to better understand the information obtained.
     For example, in a financial advisory scenario, the content returned by the data query tool may appear in English or the format is not suitable for the language model to understand. At this time, if the tool is directly exported parameters for the next step, it is very likely to cause misunderstanding. Therefore, the data can be formatted in the tool's post-function. Resp.content = "Get the market. The tool returns:" + Resp.content #The tool returns the content field may be in English or the format is not suitable for the model to understand. It needs to be formatted in the post-function
   • Tip Enhancement Rules: A large model attention enhancement solution proposed by Claude Code, when dealing with complex tasks, the large model will forget instructions or plans, through prompt word enhancement, reminding the model of the task that should be done after the tool is executed. It can be configured either by prompt or by code.
     For example, in the stock analysis scenario, the user's query is < XX market >, and the tool imported parameter is required to be the stock code, then you can specify in the pre-function what service should be called for stock code acquisition

Add Skills:

1. Click + to add the required Skills. See the self-built Skill Management operation Skill.
   
2. The added skills will be displayed in the Skills area, which supports deletion operations.
   

Knowledge

Add knowledge. Click the + icon to add three types of knowledge: term base, QA question and answer base, and general knowledge base, which can be used as additional information input when the agent executes to help the agent better generate answers. Multiple knowledge bases of each type can be added. The creation of the knowledge base can refer to the enterprise knowledge engine Create a New Knowledge Base.

Reply rule

The agent is guided to generate a reply by setting rules such as template requirements for the reply through the prompt word. Support for enabling streaming output.

Memory

Memory allows users to configure their conversation rounds, short-term memory, and long-term memory on demand.

• Short-term memory: On means that the original text of the historical session and the overview of the historical session are given to the agent as context input.
  • Number of dialogue rounds: The number of context rounds recalled when intent recognition is rolled, and the context dialogue information is used as supplementary information during query rewriting and intent recognition.
• Long-term memory: Automatically retrieve records of user portraits, preferences, and events.
  • Memory: Specifies the memory currently required from the session memory.
  • Memory usage rules: support multiple selection
    • Read Memory: Allows the agent to read historical memories in this memory bank.
    • Write Global Shared Memory: Allows the agent to precipitate the content of this conversation into a global generic portrait memory/event memory.
    • Write to agent memory: Allows the agent to precipitate the content of this conversation into a memory only for the current agent. After selecting this rule, you need to specify a memory retrieval policy.

Dialog configuration

1. Opening remarks: the fixed guiding content of the opening remarks;
2. Opener problem: Configure the fixed problem of opener guide, add up to 5;
3. Dialogue guide question: closed by default, the dialogue guide question is the related question after the answer is output. Output rules can be configured by prompts;
4. Prompt: Fixed prompt that can be configured to appear before or after the answer output. It is usually displayed as a risk warning after the answer output.
5. Display the source of the question: After opening, the content obtained through the knowledge base or online search tool in the answer will display the file or website address of the corresponding content through the corner mark.

Lifecycle Dynamic Function

An open code node, which can input the rule requirements of the agent before execution, before thinking, after thinking, before planning, during planning, and after planning.

Streaming output filter function

The middleware code framework for streaming output, the framework guarantees that each chunk will pass through two of the hooks, where you can decide it "release/replace/discard", and can do cumulative state processing across chunks by custom_context.

Security policy configuration

Support for large model application firewall functionality with BytePlus to ensure model input and output content is safe, available, and trusted.

• How to obtain parameter configuration information such as AK, SK, server address, firewall ID, region, etc., please refer to the description of key parameters of large model application firewall.
• Whether to verify the request content: After opening, the request content will be verified for security compliance.
• Whether to verify LLM output: After opening, the large model output will be verified for security compliance.
• Hit risk control prompt:
  • Content generation using security services: Prompts are generated using the internal capabilities of a large model application firewall.
  • Use a custom security prompt: You need a custom prompt in the input box.

Timeout and limited viewership of

• Speed Limit Configuration: The maximum number of requests per second (or minute), the default is null to indicate no limit.
• Error code: The error code returned when the limit viewership is reached, the default is 429.
• Error message: The prompt message returned when the limit viewership is reached, the default is "There are too many visitors at present, please try again later".
• Sub-agent timeout time: Set the timeout time of calling the downstream sub-agent, in seconds.

Basic information configuration

Click the Edit icon on the right side of My Agent in the upper left corner of the page to edit the agent's name, description, and avatar.

• Application name: Required, can be customized naming, character limit of 50.
• Application description: Required, the ability description of the agent, it is recommended to describe the ability range of the agent as clearly as possible. In the future, if other autonomous planning agents or intelligent intent planning engines schedule the agent, the description will be used as the capability positioning for planning selection, with a character limit of 500.
• Avatar: Generated by the system by default, you can also click to randomly select to switch the new avatar.

Model configuration

• Support turning on multi-modal recognition capability, after turning on, it will support multi-modal input of pictures.
• Planning Model: Select the specified model for autonomously planning execution steps based on user input intentions.
• Planning prompt word: the agent's planning prompt word, it is recommended to fill in the agent's role, task, execution requirements, restrictions and other rules.

Tool

Support adding MCP tools (tools under all MCP Servers registered in MCP), agent tools (external agents or self-built agents in the agent center), SKills (published Skills in the Skill Center).

Add an MCP tool or agent tool:

1. Click +, you can choose to add MCP tools or agent tools, about the creation of self-built MCP tools can be seen MCP.
   
2. Click the Edit button with the added MCP or agent card to edit the capability description of the MCP tool or agent under the current agent. The modified capability description is only used by the agent and will not affect the description configuration of the tools and agents.
   
3. Click the Parameter Configuration button where the MCP or agent card has been added to configure the [Pre-Execution Function], [Post-Execution Function] and [Tip Word Enhancement Rules] executed by the tool.
   
   • Pre-execution rules: Some rule actions performed before the tool is executed, often used for data conversion or processing.
     For example, in the stock analysis scenario, the user's query is < XX market >, and the tool imported parameter is required to be the stock code, then you can specify in the pre-function what service should be called for stock code acquisition
   • Post-execution rules: Some rule actions performed after the tool executes the information acquisition are often used to do information format processing, which is convenient for large models to better understand the information obtained.
     For example, in a financial advisory scenario, the content returned by the data query tool may appear in English or the format is not suitable for the language model to understand. At this time, if the tool is directly exported parameters for the next step, it is very likely to cause misunderstanding. Therefore, the data can be formatted in the tool's post-function. Resp.content = "Get the market. The tool returns:" + Resp.content #The tool returns the content field may be in English or the format is not suitable for the model to understand. It needs to be formatted in the post-function
   • Tip Enhancement Rules: A large model attention enhancement solution proposed by Claude Code, when dealing with complex tasks, the large model will forget instructions or plans, through prompt word enhancement, reminding the model of the task that should be done after the tool is executed. It can be configured either by prompt or by code.
     For example, in the stock analysis scenario, the user's query is < XX market >, and the tool imported parameter is required to be a stock code, then you can specify in the pre-function what service should be called for stock code acquisition

Add Skills:

1. Click + to add the required Skills. See the self-built Skill Management operation Skill.
   
2. The added skills will be displayed in the Skills area, which supports deletion operations.
   

Knowledge

Add knowledge. Click the + icon to add three types of knowledge: term base, QA question and answer base, and general knowledge base, which can be used as additional information input when the agent executes to help the agent better generate answers. Multiple knowledge bases of each type can be added. The creation of the knowledge base can refer to the enterprise knowledge engine Create a New Knowledge Base.

Memory

Memory allows users to configure their conversation rounds, short-term memory, and long-term memory on demand.

• Short-term memory: On means that the original text of the historical session and the overview of the historical session are given to the agent as context input.
  • Number of dialogue rounds: The number of context rounds recalled when intent recognition is rolled, and the context dialogue information is used as supplementary information during query rewriting and intent recognition.
• Long-term memory: Automatically retrieve records of user portraits, preferences, and events.
  • Memory: Specifies the memory currently required from the session memory.
  • Memory usage rules: support multiple selection
    • Read Memory: Allows the agent to read historical memories in this memory bank.
    • Write Global Shared Memory: Allows the agent to precipitate the content of this conversation into a global generic portrait memory/event memory.
    • Write to agent memory: Allows the agent to precipitate the content of this conversation into a memory only for the current agent. After selecting this rule, you need to specify a memory retrieval policy.

Dialog configuration

1. Opening remarks: the fixed guide content of the opening remarks;
2. Opener question: Configure the fixed question of opener guide, add up to 5.

Lifecycle Dynamic Function

An open code node, which can input the rule requirements of the agent before execution, before thinking, after thinking, before planning, during planning, and after planning

Streaming output filter function

The middleware code framework for streaming output, the framework guarantees that each chunk will pass through two of the hooks, where you can decide it "release/replace/discard", and can do cumulative state processing across chunks by custom_context .

Security policy configuration

Support for large model application firewall functionality with BytePlus to ensure model input and output content is safe, available, and trusted.

• How to obtain parameter configuration information such as AK, SK, server address, firewall ID, region, etc., please refer to the description of key parameters of large model application firewall.
• Whether to verify the request content: After opening, the request content will be verified for security compliance.
• Whether to verify LLM output: After opening, the large model output will be verified for security compliance.
• Hit risk control prompt:
  • Content generation using security services: Prompts are generated using the internal capabilities of a large model application firewall.
  • Use a custom security prompt: You need a custom prompt in the input box.

Timeout and limited viewership of

• Speed Limit Configuration: The maximum number of requests per second (or minute), the default is null to indicate no limit.
• Error code: The error code returned when the limit viewership is reached, the default is 429.
• Error message: The prompt message returned when the limit viewership is reached, the default is "There are too many visitors at present, please try again later".
• Sub-agent timeout time: Set the timeout time of calling the downstream sub-agent, in seconds.