Zachary Blumenfeld of Neo4j demonstrates graph-shaped agent procedures in this 20-minute AI Engineer talk.

Three graphs, three jobs

  • Execution graphs route work between agents, tools, and decision steps inside the software running an agent.
  • Context graphs connect conversation history, domain entities, and past decisions to their supporting data.
  • Procedural graphs encode reusable instructions, scripts, validation, and escalation paths.

Compile what worked into a procedure

  • Start with a written skill or successful decision traces, then compile the steps into a structured graph.
  • Package that graph with scripts, policy documents, and supporting assets in a portable skill.
  • Keep links to the original evidence so people can inspect decisions and revise procedures over time.
  • Neo4j’s Agent Memory Service is the reference architecture shown; some surrounding tools are experimental.

The loan-approval example

  • A bakery’s business-plan PDF connects to transaction records, lending policies, and earlier decisions.
  • An unconstrained agent can reach the right answer through inconsistent routes, with uncalibrated confidence.
  • The compiled version checks hard policy rules, calls a decision model when needed, and routes cases for human review.
  • Report generation and checking become explicit steps, invoked through a small command-line wrapper.

Inspect branches and measure the result

  • Persist step-level runs alongside the data, including approvals, denials, and branches taken or left unexplored.
  • Blumenfeld reports faster, cheaper runs and the ability to use a smaller model after moving routing into scripts.
  • Caveat: this is a vendor demonstration with preloaded examples and network delays. Its timing and cost remarks do not provide a controlled benchmark.

“Instead of the agent deciding all the routing steps, it’s just done programmatically.” — Zachary Blumenfeld, 15:54