HybridReActAgentStrategy

HybridReActAgentStrategy is a specialized agent strategy optimized for Mixture of Experts (MoE) architecture models.

This strategy addresses the ambiguity in internal state machines of certain MoE models when distinguishing between Tool and Completion identifiers. Unlike traditional toolchain approaches that rely on explicit ToolCall-ToolResult interactions, this hybrid approach uses ToolCall triggering combined with appending pure text directly to the context.

Inheritance

• Extends: BaseReActAgentStrategy
• Category: "agent-mixed"

Key Characteristics

• ToolCall Triggering: Initiates tool execution through standard ToolCall mechanisms
• Context-Based Integration: Appends tool results as plain text messages rather than structured ToolResult objects
• XML Tag Format: Uses <TOOL_CALL> and <TOOL_RESULT> XML tags to represent tool interactions
• MoE-Specific Optimization: Resolves issues where MoE models struggle to differentiate between tool invocation states and completion states

Properties

• regexes (ClassVar[list[tuple[re.Pattern, str]]]): Regular expressions for XML tag sanitization
• _tool_call_jinja2 (Template): Jinja2 template for rendering tool calls and results
• _process_message (list[str]): Temporary storage for processed tool messages

Constructor Parameters

• ctx (StrategyContext): Strategy context containing chat_object, configuration, and message context

Tool Function Schema

<!-- Tool Call -->
<TOOL_CALL name="tool">
    <PARAMS>
        <!-- Parameters are passed as key-value pairs -->
        <PARAM name="param1">value1</PARAM>
    </PARAMS>
</TOOL_CALL>

<!-- Tool Result -->
<TOOL_RESULT name="tool">
   Tool execution result content
</TOOL_RESULT>

Security Considerations

⚠️ Important Security Warning:

• Prompt Injection Risk: Appending tool results as plain user messages may expose the model to injection attacks if tool outputs are untrusted or unsanitized
• Minimal Sanitization: This strategy only provides basic tag pair escaping and does NOT perform semantic-level filtering or content validation
• Security Responsibility: Users MUST implement comprehensive input validation, semantic analysis, and content sanitization for tool results in production environments

Usage Example

import asyncio
from amrita_core import create_agent, minimal_init
from amrita_core.builtins.agent import HybridReActAgentStrategy


async def use_hybrid_strategy():
    # Initialize AmritaCore
    await minimal_init()

    # Create agent with hybrid strategy for MoE models
    agent = create_agent(
        url="https://api.moemodel.com",
        key="your-api-key",
        strategy=HybridReActAgentStrategy,
    )

    # Use the agent
    chat = agent.get_chatobject("Analyze this data using available tools")
    async with chat.begin():
        response = await chat.full_response()
        await chat  # Wait for the task to finish before exiting

When to Use

Use HybridReActAgentStrategy when working with:

• Mixture of Experts (MoE) models like Mixtral, Qwen-MoE, etc.
• Models that exhibit inconsistent behavior with standard ToolCall-ToolResult message pairs
• Scenarios where the model's internal state machine has difficulty distinguishing between tool invocation and completion states

When NOT to Use

Avoid HybridReActAgentStrategy when:

• Working with standard LLM providers (OpenAI, Anthropic, etc.) - use ReActAgentStrategy instead
• Security is a primary concern and you cannot implement proper input validation
• You need strict OpenAI-compatible message formatting