Hey LangGraph folks,
When you have agents delegating to sub-agents or calling tools in a complex graph, it quickly gets fuzzy:
Which human originally approved this action, and with what scope?
I built HDP (Human Delegation Provenance) as a lightweight fix. Currently in IETF draft, research backed.
A human signs a short Ed25519 token once. Agents then cryptographically extend the chain on every handoff or tool call. Verification is completely offline, no extra services, no latency.
As of April 2026:
Python + LangChain/LangGraph support is ready today via the middleware package:
https://github.com/Helixar-AI/HDP/tree/main/packages/hdp-langchain
It includes a clean HdpCallbackHandler and @hdp_node() decorator that auto-extends the chain on node transitions. Super lightweight and drop-in.
It’s already seeing real use: Some folks in the LangChain community are running it in actual scenarios today.
HDP is also now included as an app in the official Google Gemma 4 cookbook.
Would baking lighter native support for HDP into LangGraph make sense? It would give clean, tamper-evident provenance that plays nicely with LangSmith tracing and human-in-the-loop patterns, without changing how you build graphs today.
Curious if others are hitting this gap in production, and what the team thinks.
Thanks!
Siri (Helixar AI)
PS: Reposting here as originally posted in a wrong thread in talkshop