Per-step tool allowlists in LangGraph

[Mathharo]

I am building multi-tenant agent, and each tenant turn on different tools/scopes. I need allowlist tools per step and sometimes give more access after user consent, but I don’t want break cache. Any recommended approaches in LangGraph?

[augustocesarperin]

Don’t rebuild the graph. Add a policy node that writes allowed_tools to state each step
& bind the agent only to those tools (llm.bind_tools(enabled)). After consent, update state.
For cache, just include the current tool list in your cache key
(This is what’s worked for me so far)

[Mathharo]

Ok, do you usually interrupt() and resume with updated scopes, or just write consent?

[augustocesarperin]

Interrupt for consent flows, state updates for the rest. Just rebind tools from state each call to avoid stale bindings