Files
ai-agent-book/chapter2/system-hint/NOTES.md
T
liqiang b119135836
Build latest book artifacts / build (push) Canceled after 0s
dependency resolution / resolve (3.11) (push) Canceled after 0s
dependency resolution / resolve (3.13) (push) Canceled after 0s
deploy-pages / build (push) Canceled after 0s
deploy-pages / deploy (push) Canceled after 0s
i18n consistency check / check (push) Canceled after 0s
provider adoption tests / test (chapter2/context-compression) (push) Canceled after 0s
provider adoption tests / test (chapter2/prompt-injection) (push) Canceled after 0s
provider adoption tests / test (chapter2/system-hint) (push) Canceled after 0s
provider adoption tests / test (chapter3/log-sanitization) (push) Canceled after 0s
web-search-agent tests / test (push) Canceled after 0s
web-search-agent tests / agentbook (push) Canceled after 0s
ai-agent-book 精选快照(<2MB 代码与文档,来自 github.com/bojieli/ai-agent-book)
2026-08-20 13:12:50 +00:00

4.4 KiB

System-Hint Agent Implementation Notes

Comparison with Week1/Context Pattern

This project follows the same ReAct loop pattern as week1/context with the following enhancements:

Similarities to Week1/Context:

  1. ReAct Loop: Standard Reasoning + Acting pattern
  2. Command-Line Interface: Uses argparse for CLI arguments
  3. Interactive Mode: Default mode for user interaction
  4. Task Execution: execute_task() method with max iterations
  5. Kimi K3 Model: Uses the same LLM provider setup

Key Enhancements:

1. System Prompt Architecture

  • Week1/Context: Basic system prompt with tool descriptions
  • System-Hint: Enhanced system prompt with:
    • TODO list management rules
    • Error handling guidelines
    • Loop prevention strategies
    • Behavioral instructions

2. Context Management

  • Week1/Context: Manages conversation history with optional context modes
  • System-Hint: Dynamic system hints that update after each interaction:
    • Current timestamp
    • System state (directory, OS, shell)
    • TODO list status
    • Tool call counters

3. Tool Feedback

  • Week1/Context: Standard tool results
  • System-Hint: Enhanced tool results with:
    • Timestamps on each result
    • Call numbers (e.g., "Tool call #3")
    • Detailed error messages with suggestions
    • Execution duration tracking

4. Task Management

  • Week1/Context: Single-task execution
  • System-Hint: Built-in TODO list system:
    • Automatic creation for complex tasks
    • Status tracking (pending, in_progress, completed, cancelled)
    • Persistent across conversation turns

Sample Task

The default sample task demonstrates analyzing week1 and week2 projects, similar to the context project's financial analysis tasks but focused on code exploration:

# Sample task that exercises multiple tools
task = """Analyze and summarize the AI Agent projects in week1 and week2 directories:
1. Navigate to the parent directory to access both week1 and week2 folders
2. For week1 directory:
   - List all project folders
   - Read key files from projects
   - Identify the key concepts
3. For week2 directory:
   - List all project folders  
   - Read README files
   - Understand advanced features
4. Create a comprehensive analysis file
"""

Command-Line Usage

Following week1/context pattern with additional options:

# Interactive mode (default)
python main.py

# Single task execution (like week1/context)
python main.py --mode single --task "Your task here"

# Sample task (new)
python main.py --mode sample

# Feature flags (new)
python main.py --no-todo --no-timestamps --mode single --task "Simple task"

Configuration Flexibility

Unlike week1/context which has fixed context modes, system-hint allows granular control:

# Week1/Context approach
context_mode = ContextMode.FULL  # or NO_HISTORY, NO_REASONING, etc.

# System-Hint approach
config = SystemHintConfig(
    enable_timestamps=True,     # Toggle individually
    enable_tool_counter=True,
    enable_todo_list=True,
    enable_detailed_errors=True,
    enable_system_state=True
)

Best Practices Demonstrated

  1. Prevent Infinite Loops: Tool call counter shows "Tool call #N" to help agent recognize repetitive behavior
  2. Temporal Awareness: Timestamps help agent understand event sequences
  3. Task Organization: TODO lists for complex multi-step objectives
  4. Error Recovery: Detailed error messages with actionable suggestions
  5. Context Preservation: System state tracking across tool calls

Testing

Similar to week1/context with additional component tests:

# Basic component tests
python test_basic.py

# Quick demonstration
python quickstart.py

# Full interactive testing
python main.py

Key Learnings

  1. System hints significantly improve agent efficiency - Agents complete tasks with fewer iterations
  2. TODO lists provide structure - Complex tasks become manageable
  3. Tool counters prevent loops - Agents recognize and avoid repetitive behavior
  4. Detailed errors enable recovery - Agents can adapt strategies based on specific error information
  5. Timestamps provide context - Useful for multi-session or long-running tasks

Future Enhancements

Potential improvements building on this foundation:

  • Memory persistence across sessions
  • Collaborative TODO lists for multi-agent systems
  • Adaptive hint generation based on task complexity
  • Performance metrics tracking
  • Integration with external task management systems