ArbiterOS-alpha¶

Policy-driven governance layer for LangGraph

ArbiterOS-alpha is a lightweight governance framework that wraps LangGraph, enabling policy-based validation and dynamic routing without modifying the underlying graph structure.

Key Features¶

🔒 Policy-Driven Execution: Validate execution constraints before and after instruction execution
🔀 Dynamic Routing: Route execution flow based on policy conditions
📊 Evaluation & Feedback: Assess node quality with non-blocking evaluators (RL-style rewards)
📈 Execution History: Track all instruction executions with timestamps and I/O
🎯 LangGraph-Native: Minimal migration cost from existing LangGraph code
🧩 Decorator-Based: Use @instruction decorator for lightweight governance
🔓 Zero Lock-In: Remove ArbiterOS by removing decorators and policies

Quick Example¶

from arbiteros_alpha import ArbiterOSAlpha, ThresholdEvaluator
from arbiteros_alpha.policy import HistoryPolicyChecker, MetricThresholdPolicyRouter
from arbiteros_alpha.instructions import CognitiveCore

# Create ArbiterOS instance
os = ArbiterOSAlpha()

# Add policy checker (pre-execution validation)
os.add_policy_checker(
    HistoryPolicyChecker(
        name="no_direct_toolcall",
        bad_sequence=["generate", "toolcall"]
    )
)

# Add policy router (post-execution routing)
os.add_policy_router(
    MetricThresholdPolicyRouter(
        name="regenerate_on_low_confidence",
        key="confidence",
        threshold=0.6,
        target="generate"
    )
)

# Add evaluator (quality assessment)
os.add_evaluator(
    ThresholdEvaluator(
        name="confidence_check",
        key="confidence",
        threshold=0.7,
        target_instructions=[CognitiveCore.GENERATE]
    )
)

# Decorate your functions
@os.instruction(CognitiveCore.GENERATE)
def generate(state):
    return {"response": "AI response", "confidence": 0.85}

@os.instruction(CognitiveCore.EVALUATE)
def evaluate(state):
    return {"confidence": 0.8}

Design Philosophy¶

ArbiterOS wraps LangGraph, not replaces it.

Minimal Migration Cost: Existing LangGraph code works with minimal changes
Native LangGraph API: Standard add_node(), add_edge(), compile(), invoke(), stream()
Governance is Additive: Not replacing core functionality

Getting Started¶

Core Concepts¶

Policy Architecture - PolicyChecker (pre-execution) and PolicyRouter (post-execution)
Evaluators - Quality assessment and RL-style reward signals

Documentation Structure¶

Getting Started: Installation and quick start guides
Concepts: Core architectural concepts and design patterns
API Reference: Auto-generated API documentation from code
Examples: Practical examples and use cases
Development: Contributing guidelines and testing documentation