LangChain Integration

Auto-capture all agent activity through LangChain's callback system.

Installation

pip install agent-forensics[langchain]

Requires langchain-core >= 0.3.

Usage

from agent_forensics import Forensics
from langchain_ollama import ChatOllama
from langgraph.prebuilt import create_react_agent

f = Forensics(session="langchain-demo", agent="react-agent")

# Create your agent as usual
llm = ChatOllama(model="mistral")
agent = create_react_agent(llm, tools=[...])

# Add forensics with one line
result = agent.invoke(
    {"messages": [("user", "Find me a wireless mouse")]},
    config={"callbacks": [f.langchain()]},
)

# Generate report
print(f.report())

What Gets Captured

Event	Source	Details
llm_call_start/end	on_chat_model_start / on_chat_model_end	Full message history, model response
tool_call_start/end	on_tool_start / on_tool_end	Tool name, input, output
decision	on_agent_action	Agent's chosen action + reasoning
final_decision	on_agent_finish	Final output
error	on_llm_error / on_tool_error	Error details
prompt_drift	Automatic	System prompt changes between steps

Prompt Drift Detection

The LangChain integration automatically tracks system prompt changes. If the system message in the LLM call differs from the previous call, a prompt_drift event is recorded.

# No manual prompt_state() calls needed!
# The callback handler extracts system messages from each LLM call

Full Example

from agent_forensics import Forensics
from langchain_ollama import ChatOllama
from langchain_core.tools import tool
from langgraph.prebuilt import create_react_agent

@tool
def search_products(query: str) -> str:
    """Search the product catalog."""
    return '{"results": [{"name": "Logitech M750", "price": 45}]}'

@tool
def purchase(product: str, quantity: int) -> str:
    """Purchase a product."""
    return f'{{"status": "confirmed", "product": "{product}", "qty": {quantity}}}'

f = Forensics(session="shopping-session", agent="shopping-agent")

llm = ChatOllama(model="mistral")
agent = create_react_agent(llm, tools=[search_products, purchase])

result = agent.invoke(
    {"messages": [("user", "Buy the cheapest wireless mouse")]},
    config={"callbacks": [f.langchain()]},
)

# Forensic analysis
print(f"Events: {len(f.events())}")
print(f.report())

failures = f.classify()
for fail in failures:
    print(f"[{fail['severity']}] {fail['type']}: {fail['description']}")