Skip to content

Single-Turn Agent CoT (agent_cot_tools)

Single-turn agent Chain of Thought datasets focus on training models to handle complete tasks requiring tool usage in a single interaction. This format captures the full reasoning process from initial analysis through tool execution to final synthesis.

When to Use Single-Turn Agent CoT

Single-turn agent CoT is ideal for: - Complete task resolution where the user's request can be fully addressed in one response - Tool reasoning training where you want models to learn systematic tool selection - Parameter construction training for building accurate function calls - Result synthesis where multiple tool outputs need to be combined into coherent answers

Format Structure

The AgentCoTWithTools schema captures six key components:

{
  "question": "User's question or request",
  "available_tools": [list of tool definitions],
  "initial_analysis": "Understanding of what's needed",
  "tool_planning": [reasoning steps for tool selection],
  "tool_executions": [actual tool calls and results],
  "result_synthesis": "How results combine to answer question",
  "final_answer": "Complete response to user"
}

Detailed Schema Breakdown

question (string)

The user's original question or request that requires tool usage.

Examples: - "What's the weather in London and Paris?" - "Calculate the tip for a $45.67 dinner and find a good restaurant nearby" - "Book a table for 4 at an Italian restaurant tomorrow at 7 PM"

available_tools (list)

Complete tool definitions using Pydantic schemas. These are the tools the agent can choose from.

"available_tools": [
  {
    "name": "get_weather",
    "description": "Get current weather conditions for a location",
    "parameters": [
      {
        "name": "location",
        "type": "str",
        "description": "City name or location",
        "required": true
      }
    ],
    "returns": "Weather data including temperature and conditions",
    "category": "information"
  }
]

initial_analysis (string)

The agent's understanding of what the user needs and what information or actions are required.

Examples: - "User wants weather information for two cities. I need to retrieve current weather data for both London and Paris and present a comparison." - "This requires two tasks: calculating an 18-20% tip on $45.67, and finding restaurant recommendations near the user's location."

tool_planning (list)

Step-by-step reasoning for tool selection and parameter construction. Each step includes:

{
  "step_number": 1,
  "reasoning": "Why this tool is needed at this point",
  "selected_tool": {tool definition},
  "parameter_reasoning": {
    "param_name": "explanation of how this parameter value was determined"
  },
  "expected_result": "What the tool should return and how it helps"
}

Example: 

{
  "step_number": 1,
  "reasoning": "Need current weather data for London as specified in the user's request",
  "selected_tool": {
    "name": "get_weather",
    "description": "Get current weather conditions for a location",
    // ... full tool definition
  },
  "parameter_reasoning": {
    "location": "User explicitly mentioned London as first city"
  },
  "expected_result": "Current weather conditions for London including temperature, conditions, and other relevant data"
}

tool_executions (list)

Actual function calls made with their results. Each execution includes:

{
  "function": "tool_name",
  "arguments": {parameter dictionary},
  "reasoning": "Brief explanation of why executing now",
  "result": "The actual result returned from the tool"
}

Example: 

{
  "function": "get_weather",
  "arguments": {"location": "London"},
  "reasoning": "Executing weather lookup for London as planned",
  "result": "London: 15°C, overcast, 80% humidity, light rain expected"
}

result_synthesis (string)

Explanation of how the tool results are combined and interpreted to address the original question.

Example: - "Both weather reports show cooler temperatures with London having rain while Paris is clearer. Combined this data to provide a comprehensive comparison for the user's planning purposes."

final_answer (string)

The complete, natural response to the user incorporating all tool results.

Example: - "Here's the current weather for both cities: London is 15°C with overcast skies and light rain expected (80% humidity), while Paris is warmer at 18°C with partly cloudy conditions (65% humidity). Paris has better weather today if you're planning outdoor activities."

Configuration Example

YAML Configuration

dataset_system_prompt: "You are an AI assistant that excels at systematic tool usage and reasoning."

data_engine:
  generation_system_prompt: "Generate realistic scenarios requiring tool usage with detailed reasoning traces."
  provider: "openai"
  model: "gpt-4o-mini"
  conversation_type: "agent_cot_tools"
  available_tools:
    - "get_weather"
    - "search_web"
    - "calculator"
    - "book_restaurant"
  max_tools_per_query: 3

dataset:
  creation:
    num_steps: 20
    batch_size: 5
    sys_msg: false  # Agent format doesn't use system messages
  save_as: "single_turn_agent.jsonl"

Python API

from deepfabric import DataSetGenerator
from deepfabric.tools.defaults import get_default_tools

engine = DataSetGenerator(
    generation_system_prompt="You excel at tool reasoning and selection.",
    provider="openai",
    model_name="gpt-4o-mini",
    conversation_type="agent_cot_tools",
    available_tools=["get_weather", "search_web", "calculator"],
    max_tools_per_query=2,
)

dataset = engine.create_data(
    num_steps=10,
    batch_size=2,
    sys_msg=False
)

Best Practices

Tool Planning Quality

  • Be specific about why each tool is selected
  • Explain parameter construction based on user input
  • Set clear expectations for what each tool should return

Tool Execution Realism

  • Match planning to execution - tools used should align with the plan
  • Provide realistic results that tools would actually return
  • Include error cases where tools might not have complete information

Result Synthesis

  • Address the original question directly
  • Integrate all tool results meaningfully
  • Provide actionable information when possible

Common Use Cases

Weather and Location Services

{
  "question": "What's the weather like in Tokyo and should I bring an umbrella?",
  "tool_planning": [
    {
      "reasoning": "Need current weather conditions for Tokyo to assess umbrella necessity",
      "selected_tool": {...weather_tool...},
      "parameter_reasoning": {"location": "Tokyo specified by user"},
      "expected_result": "Current weather including precipitation probability"
    }
  ]
}

Calculation and Information Lookup

{
  "question": "If I invest $5000 at 7% annual interest for 3 years, how much will I have?",
  "tool_planning": [
    {
      "reasoning": "Need to calculate compound interest with given parameters",
      "selected_tool": {...calculator_tool...},
      "parameter_reasoning": {
        "expression": "Compound interest formula: P(1+r)^t where P=5000, r=0.07, t=3"
      },
      "expected_result": "Final investment value after 3 years"
    }
  ]
}

Multi-Tool Workflows

{
  "question": "Find a good Italian restaurant in NYC and calculate the tip for a $85 dinner",
  "tool_planning": [
    {
      "step_number": 1,
      "reasoning": "Need to search for Italian restaurants in New York City",
      "selected_tool": {...search_tool...}
    },
    {
      "step_number": 2,
      "reasoning": "Need to calculate appropriate tip amount for $85 bill",
      "selected_tool": {...calculator_tool...}
    }
  ]
}

Training Benefits

Models trained on single-turn agent CoT data demonstrate: - Improved tool selection accuracy across different domains - Better parameter construction with fewer errors - Enhanced reasoning transparency that users can follow - More efficient tool usage with reduced unnecessary calls - Better error handling when tools return unexpected results

This format creates training data that teaches models not just to use tools, but to think systematically about tool usage in a way that's both effective and explainable.