langchain-agentlattice

AgentLattice governance wrapper for LangChain tools. Gates every tool invocation through AgentLattice's audit trail, policy engine, and approval flow.

Install

pip install langchain-agentlattice

Quickstart

import os
from agentlattice import AgentLattice
from langchain_agentlattice import govern
from langchain_community.tools import DuckDuckGoSearchRun

al = AgentLattice(api_key=os.environ["AL_API_KEY"])
search = DuckDuckGoSearchRun()

# Wrap any LangChain tool — gate() runs before every invocation
governed_search = govern(search, al=al)

# Drop-in replacement — same name, description, interface
tools = [governed_search]

Handle denials:

from agentlattice import AgentLatticeDeniedError, AgentLatticeTimeoutError

try:
    result = await governed_search._arun("latest AI news")
except AgentLatticeDeniedError as e:
    print(f"Denied: {e.reason}, policy: {e.policy}")
except AgentLatticeTimeoutError as e:
    print(f"Approval timed out: {e.approval_id}")

How it works

govern() wraps any BaseTool in a GovernedTool that:

1. Calls al.gate(action_type) before executing
2. If approved → delegates to the wrapped tool's _arun() / _run()
3. If denied → raises AgentLatticeDeniedError (tool does not execute)
4. If timed out → raises AgentLatticeTimeoutError

This is a tool wrapper, not a callback handler. LangChain callbacks are notifications — they cannot block tool execution. Only the wrapper pattern preserves the gate() semantics.

Custom action types

# Default: "tool.{tool.name}"
governed = govern(my_tool, al=al)

# Override
governed = govern(my_tool, al=al, action_type="search.web.execute")