The graph shape comes from the compiled graph, the eval results from the last recorded
experiment. Nothing here is typed into the page by hand.
The graph
9 nodes, 15 edges,
10 of them conditional
Built with LangGraph. The conditional edges are the product rules: a failed payee check
routes to a stop before consent is read, an amount over the ceiling routes to a human.
The pause is interrupt(), resumed with
Command(resume=…), so the earlier checks never run twice.
Read the graph definition
loading…
The evals
13 of 13 cases passing
13 scenarios in a versioned LangSmith dataset, scored by 5 evaluators. Each case
asserts where the run ended and which nodes it must never have touched, because the
question worth asking of an agent with a payment API is whether it ever reached the node
that moves money.
The pause, and where it lives
postgres
A paused run is written to a checkpointer, so an approval arriving later resumes it in a
different process entirely. On this deployment that is Neon Postgres in Frankfurt. Run it
locally without a database and it falls back to memory, which is why the eval suite needs
no configuration.