react-agent-compensation 0.1.0

Framework-agnostic compensation/rollback library for ReAct agents

react-agent-compensation

A framework-agnostic compensation/rollback library for ReAct agents.

Features

• Framework-agnostic Core: Works with any agent framework
• LangChain Adaptor: First-class LangChain/LangGraph integration
• Compensation Patterns: Automatic rollback on failures
• Retry Strategies: Exponential backoff with jitter
• Dependency Tracking: Topological sort for correct rollback order
• MCP Integration: Auto-discover compensation pairs from tool schemas

Installation

pip install react-agent-compensation

With LangChain support:

pip install react-agent-compensation[langchain]

With LLM-based extraction:

pip install react-agent-compensation[llm]

Quick Start

from react_agent_compensation.langchain_adaptor import create_compensated_agent

agent = create_compensated_agent(
    model="gpt-4",
    tools=[book_flight, cancel_flight],
    compensation_mapping={"book_flight": "cancel_flight"},
)

result = agent.invoke({"messages": [("user", "Book me a flight to NYC")]})

Core Components

RecoveryManager

The brain of the compensation system:

from react_agent_compensation.core import RecoveryManager

manager = RecoveryManager(
    compensation_pairs={"book_flight": "cancel_flight"},
    alternative_map={"book_flight": ["book_flight_backup"]},
)

# Record before execution
record = manager.record_action("book_flight", {"dest": "NYC"})

# Mark complete on success
manager.mark_completed(record.id, result={"booking_id": "123"})

# Rollback on failure
manager.rollback()

Extraction Strategies

Multiple strategies for extracting compensation parameters:

• State Mappers: Custom lambda functions
• Schema-based: Declarative path mappings
• Heuristic: Auto-detect common ID fields
• LLM-based: Use LLM for complex extraction

License

MIT