Make Your AI Agents Predictable With Graph Engineering - Zachary Blumenfeld, Neo4j

Highlights

• Graph engineering designs agent work as an explicit graph where nodes perform work and edges route state between them.

• Compiling written procedural material into deterministic graph structures increases task reward and overall performance.

• Storing context graphs alongside organizational data adds a portability layer that works across any runtime.

• Compiling a loan approval policy into a graph-shaped skill reduced execution time from 7.5 minutes down to roughly 2 minutes while lowering costs.

• Deterministic execution scripts remove the burden of agent routing, allowing smaller models like Haiku to handle complex workloads.

Timeline

1. 00:00 - 05:00

   Defining Graph Engineering and Core Graph Types

   Graph engineering evolves from prompt and context engineering by utilizing established computer science concepts like state machines and workflow graphs. Execution graphs manage routers and agents within an harness, while context graphs track decision traces and procedural graphs capture skills and step-by-step logic. Neo4j places a knowledge layer between data sources like Databricks or Snowflake and agent runtimes to enable interoperability.

   • Graph engineering is the practice of designing agent work as an explicit graph with nodes doing work and edges routing state.
   • Three primary graph types include execution graphs, context graphs, and procedural graphs.
   • A knowledge layer sits between systems of record and agent runtimes to provide standard protocols and artifacts.
2. 05:00 - 11:00

   Context Graphs, Memory, and Compiling Skills

   Research demonstrates that converting natural language procedure files and reference assets into deterministic graph scripts allows smaller models to achieve higher task rewards. This structure maintains complete provenance back through every tool call and decision. Asset managers and compliance teams use these graph structures to automate complex audit reporting and data resolution.

   • Context graphs incorporate short-term conversation memory, long-term extracted entities, and reasoning decision traces.
   • Compiling written skill files and scripts into graph structures increases overall task reward and performance.
   • Neo4j Agent Memory Service serves as a reference architecture providing versioned provenance and governance over time.
3. 11:00 - 20:25

   Loan Approval Demo and Performance Impact

   A demonstration using a loan approval scenario for Rosa's Bakery illustrates how multi-source data is ingested into a context graph. Converting variable agent runs into a structured YAML graph skill ensures deterministic rule checks and risk score evaluations. This approach eliminates unnecessary reasoning loops, resulting in faster wall-clock times and reduced model inference expenses.

   • A loan approval process links PDF business plans with transactional data from Databricks and past lending policies.
   • Compiling inconsistent execution traces into a graph YAML shape allows programmatic script routing instead of relying on agent decisions.
   • Programmatic skill routing drops execution time from 7.5 minutes down to 2 minutes while enabling cheaper models like Haiku.