LangChain Integration Guide¶

Integrate ACGP with LangChain for automatic governance of chains, tools, and agents.

Overview¶

LangChain is a comprehensive framework for building LLM applications. ACGP integrates via callbacks to provide governance for tool calls and agent actions.

Installation¶

pip install acgp-langchain

Basic Usage¶

from langchain.agents import initialize_agent, AgentType
from langchain_openai import OpenAI
from acgp_langchain import GovernedChain
from acgp import GovernanceSteward

# Your LangChain agent
llm = OpenAI(temperature=0)
agent = initialize_agent(tools, llm, agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION)

# Add governance
steward = GovernanceSteward(blueprint_file="blueprint.yaml")
governed_agent = GovernedChain(agent, steward)

# Use normally - governance automatic!
result = governed_agent.run("Process this request")

Features¶

Automatic trace generation: Extracts reasoning from agent scratchpad
Tool call interception: Evaluates before tool execution
ReAct pattern support: Parses Thought/Action/Observation
Intervention handling: Raises exceptions for blocked actions
Callback integration: Uses LangChain's callback system

Advanced Features¶

Override Blocked Actions¶

# After human approval
result = governed_agent.run_with_override(
    trace_id="blocked_trace_id",
    "Previously blocked query"
)

Get Governance Metrics¶

metrics = governed_agent.get_metrics()
print(f"Total interventions: {metrics['total_interventions']}")

Exception Handling¶

from acgp_langchain import InterventionError

try:
    result = governed_agent.run("Risky query")
except InterventionError as e:
    print(f"Action blocked: {e.message}")
    print(f"Trace ID: {e.trace_id}")

Control Intervention Behavior¶

# Don't raise exceptions on interventions
governed_agent = GovernedChain(
    agent,
    steward,
    raise_on_intervention=False  # Log instead of raising
)

Using the Callback Directly¶

from acgp_langchain import ACGPCallback

# Use callback directly with any LangChain component
callback = ACGPCallback(steward, raise_on_intervention=True)
result = chain.run("Query", callbacks=[callback])

Examples¶

Example implementations available in the acgp-langchain package:

Governed Agent: LangChain agent with ACGP governance
Tool Interception: Automatic tool call evaluation
ReAct Pattern: Integration with ReAct agent reasoning

LangGraph Integration