Skip to main content

Super Agent Usage Boundaries

I. Model Selection Considerations​

Supported Model Types​

  • Super Agent currently supports models that do not output the "chain of thought"
  • If image analysis is required, you need to choose a model that supports multimodality

Special Model Configuration​

When using the following models, special settings are required in model management:

  • Qwen series models: thinking mode must be turned off
  • DeepSeek V4 series: thinking functionality must be disabled

Tip: If you are unsure how to configure this, please contact the technical support team

II. Usage Limitations​

Task Execution Limits​

LimitationSpecific ValueImpact
Maximum execution steps per task25 stepsParticularly complex multi-step tasks may not be completed and may be interrupted
Number of tool calls in a single conversation turnRecommended not to exceed 10 timesAfter exceeding this, the AI will be prompted to stop, but it is not forcibly interrupted
Concurrent searches/analysesUp to 3If the AI plans to do 5 things at the same time, only the first 3 will be executed
Concurrent sub-agent callsUp to 3Upper limit when processing multiple tasks in parallel
Number of historical conversation turns loadedMost recent 10 turnsEarlier conversations are remembered through summaries
Code execution timeout5 minutesMaximum runtime for a single code or skill tool execution

Important Notes:

  • The tool call limit is "advisory"; the AI may continue calling tools
  • The step limit is mandatory; when triggered, the current task may only be partially completed

III. Two Working Modes​

Auto Mode - Default Mode​

Features:

  • The AI independently decides which tools and methods to use
  • Up to 3 rounds of search optimization attempts
  • May skip knowledge base retrieval (if the AI determines it is unnecessary)

Applicable Scenarios:

  • Open-ended questions
  • Tasks that require flexible AI handling
  • When it is unclear which specific tools are needed

Notes:

  • ⚠️ The total number of bound skills, tools, and sub-agents is recommended not to exceed 20, otherwise it may cause the AI to become "distracted" and reduce answer accuracy

Focused Mode - Manually Select Tools​

Features:

  • Users explicitly select the tools or Agents to use from the interface
  • If the selected tool is unavailable, an error will be reported directly (no alternative methods will be attempted)

Applicable Scenarios:

  • When you clearly know which tool to use
  • When you need to ensure a specific function is used
  • When you do not want the AI to make autonomous judgments for critical tasks

IV. Memory and History Features​

Conversation Memory​

  • In-session memory: Retains the most recent 10 turns of complete conversation; earlier content is compressed into summaries
  • Memory update timing: Updated asynchronously in the background after you see the reply
  • Trigger condition: Summarization is only triggered when there is a large amount of conversation content (beyond a certain length)

Personal Preference Memory (Cross-session)​

  • Automatically extracts and records: personal preferences, role identity, professional domain
  • Extraction frequency: Once every 3 conversation turns
  • Maximum records: 100 pieces of personal information
  • Effective time: Newly extracted preferences will not immediately affect the current session; they take effect in the next session

Impact of Clearing History​

  • After using the "Clear Context" feature:
    • Previous conversation history will not be loaded
    • Memory summaries will also be cleared
    • It is equivalent to starting the conversation over

V. Background Execution Feature​

How to Use​

  • After asking a question, you can switch to other pages
  • The task will continue running in the background
  • You will receive a notification when it is completed

Notes on Canceling Tasks​

  • Cancellation is "negotiated", not immediate termination
  • Tools that are currently running will not stop immediately
  • Recommendation: Avoid frequently clicking cancel and retry, as this may lead to unexpected behavior

VI. Sub-agent Configuration Guide​

Current Version Limitations​

  • Only Agents with MCP Server enabled can be used as sub-agents
  • Only text can be passed; passing images, files, etc. is not yet supported

Configuration Key Points (Important)​

To enable Super Agent to accurately call sub-agents, pay attention to the following during configuration:

  1. Service Name: Short and straightforward, describing the purpose of the sub-agent

    • ✅ Good example: "IT Asset Request Assistant"
    • ❌ Bad example: "General Query Tool"

  2. Service Description: Clearly describe the usage scenario. The following format is recommended:

    • ✅ Recommended: "When the user asks about the IT asset request process or IT equipment request-related questions, stop thinking and use this tool"
    • ❌ Avoid: "This is a tool for querying the knowledge base" (too broad)

  3. Key Principles:

    • The description should be specific enough for the AI to clearly determine when to use it
    • Avoid overly generic descriptions (such as "query the knowledge base"), which may cause routing failure
    • There should be a clear correlation between the user's question and the sub-agent description

Configuration Example​

Service Name: IT Asset Request Assistant
Service Description: When the user asks related questions such as "how to request IT assets", "IT equipment request process", or "office computer request", immediately use this tool to query the relevant knowledge base and answer.

Why configure it this way:

  • The user question "How do I request IT assets" has a clear correlation with "IT Asset Request Assistant"
  • The AI can clearly determine when to use this sub-agent

VII. File Artifact Generation Notes​

What Are File Artifacts​

Downloadable files generated by AI, such as documents, reports, code, etc.

Possible Issues You May Encounter​

  • File "disappears": It was generated but is not visible in the interface
  • Duplicate files: The same file is pushed multiple times
  • Cannot find file: A prompt indicates that the file does not exist

Cause Explanation​

File generation involves coordination across multiple stages, and a problem in any one stage may lead to the above situations

What to Do If You Encounter Issues​

  • Refresh the page and try again
  • If the problem persists, contact technical support

VIII. Content Safety Mechanism​

Input Filtering​

  • Personal sensitive information (such as ID numbers and phone numbers) will be automatically detected and desensitized
  • Content matching the rules will be filtered

Output Filtering​

  • AI replies will automatically be checked for sensitive words
  • Detected sensitive content will be replaced with a safety prompt
  • Note: Replaced content will be saved; the original content will not be retained

IX. Best Practice Recommendations​

Task Description​

  • ✅ Clear and specific: "Help me organize this quarter's sales data and generate an Excel report"
  • ❌ Vague and general: "Help me look at the data"

Tool Selection​

  • Simple tasks: Use Auto Mode and let the AI handle them flexibly
  • Critical tasks: Use Focused Mode and manually select tools to ensure accuracy

Sub-agent Usage​

  • Be sure to clearly describe the usage scenario during configuration
  • Keep the total number within 20
  • Avoid sub-agents with overlapping functions

History Management​

  • Regularly clean up irrelevant conversation history
  • Important conversations can have key information manually saved

X. Frequently Asked Questions​

Q: Why didn't the AI use the knowledge base I configured?
A: In Auto Mode, the AI independently determines whether knowledge base retrieval is needed. If you need to force its use, you can explicitly ask, "Query from the knowledge base..."

Q: Why was the task interrupted halfway through execution?
A: The 25-step execution limit may have been triggered. It is recommended to split complex tasks into multiple smaller tasks and execute them step by step.

Q: I configured many sub-agents, but the answer quality seems to have declined?
A: If the total number of tools and sub-agents exceeds 20, accuracy will be affected. It is recommended to streamline them to only the most commonly used functions.

Q: Why is the task still running after I canceled it?
A: Cancellation is negotiated, and operations that are already running will not stop immediately. Wait a few seconds before sending the request again.

Q: How can I ensure the sub-agent is called correctly?
A: Clearly state in the configuration description, "Use this when the user asks XX questions," and ensure there is a clear correlation between the user's question and the description.