Reasoning Patterns API Reference¶

Overview¶

AgentiCraft provides three advanced reasoning patterns that enable agents to solve complex problems with transparency and structured thinking.

Available Patterns¶

Chain of Thought (CoT)¶

Linear, step-by-step reasoning with confidence tracking and alternative generation.

Tree of Thoughts (ToT)¶

Multi-path exploration with scoring, pruning, and optimal path selection.

ReAct ¶

Combines reasoning with actions, creating dynamic thought-action-observation cycles.

Quick Start¶

from agenticraft.agents.reasoning import ReasoningAgent

# Create agent with specific pattern
agent = ReasoningAgent(
    name="SmartAgent",
    reasoning_pattern="chain_of_thought",  # or "tree_of_thoughts", "react"
    pattern_config={
        # Pattern-specific configuration
    }
)

# Execute reasoning
response = await agent.think_and_act("Your problem here")

# Access results
print(response.content)  # Final answer
print(response.reasoning)  # Human-readable reasoning
for step in response.reasoning_steps:
    print(f"{step.number}: {step.description} (confidence: {step.confidence})")

Pattern Selection¶

Automatic Selection¶

# Let the agent choose the best pattern
agent = ReasoningAgent(name="AutoAgent")
pattern = agent.select_best_pattern("Your problem")

Manual Selection Guide¶

Problem Type	Recommended Pattern	Example
Sequential analysis	Chain of Thought	"Explain how photosynthesis works"
Creative exploration	Tree of Thoughts	"Design a mobile app for seniors"
Research & tool use	ReAct	"Find the current GDP of Japan"

Common Types¶

Base Types¶

from agenticraft.reasoning.patterns.base import (
    ReasoningStep,
    ReasoningTrace,
    PatternConfig
)

Pattern-Specific Types¶

# Chain of Thought
from agenticraft.reasoning.patterns.chain_of_thought import ThoughtStep

# Tree of Thoughts
from agenticraft.reasoning.patterns.tree_of_thoughts import (
    TreeNode,
    NodeStatus,
    NodeType
)

# ReAct
from agenticraft.reasoning.patterns.react import (
    StepType,
    ReactStep
)

Integration with Agents¶

All reasoning patterns integrate seamlessly with ReasoningAgent:

# The agent handles pattern initialization and execution
agent = ReasoningAgent(
    reasoning_pattern="tree_of_thoughts",
    pattern_config={
        "max_depth": 4,
        "beam_width": 3
    }
)

# Pattern is used automatically
response = await agent.think_and_act("Design a logo")

Performance Metrics¶

Pattern	Complexity	Time (simple)	Time (complex)	Memory
CoT	O(n)	~50ms	~150ms	Low
ToT	O(b^d)	~200ms	~500ms	High
ReAct	O(n)	~100ms	~300ms+	Medium