How we add runtime security to LangChain agents in production

Talking Shop
1

Been seeing a lot of questions about production safety for LangChain agents so wanted to share what we use.

The problem: once a LangChain agent has tool access in production, there’s no built-in way to intercept and block dangerous actions before they execute.

Here’s a lightweight approach using Vaultak:

pip install vaultak

from vaultak import Vaultak

vt = Vaultak(api_key="vtk_...")

with vt.monitor("my-agent"):
    result = agent.invoke({"input": user_input})

Every tool call now gets risk-scored before execution. You can add policies to block specific actions:

vt.policy.create({
    "name": "no-prod-deletes",
    "action": "delete",
    "resource": "production_*",
    "effect": "deny",
    "priority": 1
})

There’s also a free scanner at AI Agent Risk Scanner, Vaultak if you want to check your agent’s risk profile without writing any code first.

Full guide with more examples:

Happy to answer questions about the implementation.

2

@samueloladji-beep , why not wrap this in a middleware, rather than using it in a context manager.

3

middleware is actually cleaner for production use. Here’s how you’d do it:

from langchain.callbacks.base import BaseCallbackHandler
from vaultak import Vaultak

class VaultakHandler(BaseCallbackHandler):
def init(self, api_key: str, agent_name: str):
self.vt = Vaultak(api_key=api_key)
self.agent_name = agent_name

def on_tool_start(self, tool, input_str, **kwargs):
    self.vt.log_action(
        agent_id=self.agent_name,
        action_type="tool_call",
        resource=tool["name"],
        payload=input_str
    )

agent.invoke(
{“input”: user_input},
config={“callbacks”: [VaultakHandler(api_key=“vtk_…”, agent_name=“my-agent”)]}
)

The context manager approach is simpler for quick setup. The callback handler is better for production; it hooks directly into LangChain’s execution lifecycle so you get per-tool visibility rather than wrapping the whole invocation.
We’re actually adding a native LangChain integration to the SDK. Will post here when it’s ready.