For Neo4j’s earlier visit with Chief Scientist Dr. Jim Webber on graphs as LLM context, see the S2E4 archive summary.
Most enterprises are racing to deploy agents. Sudhir Hasbe, President and Chief Product Officer at Neo4j, argues the harder shift is upstream: software is moving from UI shells where humans enter data to agentic systems that need knowledge—structured, relational, bounded—not another document dump.
On a recent Invisible Machines episode with Josh Tyson and Robb Wilson, Hasbe meets the show where it has been heading all season: organizational AGI is not wide AGI. It is scoped intelligence—knowing the right things before doing anything at all.
Know Before You Do
Robb Wilson opens with the onboarding metaphor that frames the episode. When you hire someone on day one, you do not hand them a scalpel and say start carving; you teach them what they need to know. Agents fail for the same reason most AI projects fail in enterprises: data is not structured for reasoning systems. Gartner talking about knowledge graphs is, in Hasbe’s dry read, a lagging indicator that customers are already living the problem.
The shift is from collecting data to making knowledge available—so autonomous agents can decide with context rather than confabulate from similarity search. Graph databases enter not as a trendy storage format but as the natural representation for relationships that keep growing: customer to product to channel to market behavior, each edge a different kind of fact that tables flatten into noise.
Feature Reduction, Not More Context
Wilson names the design discipline the industry under-sells: feature reduction. Agents need the wide knowledge of the organization compressed to the relevant slice for this task—not a million tokens when someone asks how their day is going.
Hasbe maps this to the old machine-learning world of feature stores: in agentic systems, the analogue is the context store. A support agent should see who the customer is, what they bought, and what matters about the product—not the entire supply chain graph unless the question demands it. Too much irrelevant context burns tokens and increases hallucination—the same overfitting failure mode, now priced per request.
Wilson pushes further: organizational AGI is about boundaries, not breadth. The goal is not more knowledge in the world but exactly enough knowledge to complete the task— not too much, not too little. That is where practitioners struggle, and where graph semantics help humans and agents alike reason over connected facts instead of isolated cells.
Data, Information, Knowledge, Intelligence, Wisdom
The episode’s conceptual spine is Wilson’s complexity continuum—a number becomes a date becomes a birthday becomes wisdom (too young to drink, old enough to vote). Agents tasked with real work cannot be fed the number; they need the wisdom layer.
Hasbe extends the ladder. Graphs emphasize relationships over points: a purchase is not a row—it is a dated edge between customer and product with attributes attached. As those relationships accumulate, information becomes knowledge. The next step is intelligence—decisions orchestrated on top of that knowledge. Wisdom, for Hasbe, is organizational memory: decision traces, what agents and humans did last time, what worked, what failed, what must be auditable next time.
A goldfish agent that forgets every session cannot improve. Claims processing that ignores how humans decided similar cases yesterday will repeat mistakes at machine speed. Wisdom lives in the loop between action and recorded outcome—and graph storage is where Hasbe sees ontologies, operational data, and memory converging.
Three Layers: Ontology, Data, Memory
Hasbe describes Neo4j’s architectural bet in three parts. First, ontologies and metadata—a semantic map of what lives where across Snowflake, Databricks, Salesforce, and the rest, including which systems make which decisions. Second, the data itself—native graph storage for performance or virtual graphs that leave source systems in place while presenting a unified relationship surface. Third, memory and context—the continuous learning substrate that turns a one-shot Q&A bot into something that accumulates organizational judgment.
Wilson connects this to agent sprawl. An agent registry sounds simple until you need agents to cross silos, find capabilities, and stop caring where bytes physically live—as long as the graph knows how to fetch them. People call graph databases complex; Wilson argues the alternative—twenty systems with no map—is the truly unmanageable architecture.
Cost, Truth, and Siloed Agents
Josh Tyson brings the CFO lens: Cassie Kozyrkov and Brian Katanzaro have both noted that AI compute can exceed employee cost if agents spin in circles. Hasbe agrees—governed, deterministic answers should stay short; open-ended reasoning can afford richer context. The failure mode is building isolated agents that each maintain their own partial worldview while claiming to serve the same customer.
Organizational AGI requires shared artifacts: data can live anywhere, but the graph tells every agent which truth to reference. Without that, support, recommendations, and fulfillment agents optimize locally and contradict globally.
Live Demo: Learning Machine on Neo4j
The episode’s second half is a live walkthrough of UX Magazine’s Learning Machine—a guided education agent built on a Neo4j knowledge model of Roger Schwab’s lean knowledge management book (developed after NASA’s Columbia disaster).
Josh shows how chapters become tagged canonical ideas linked as nodes; Robb describes flipping the empty chat window into use case zero: assess what the learner already knows, light up the graph, compute a personalized path (traveling-salesman style) through concepts, and teach—rather than waiting for a prompt. The demo extends to a Growth Hub career journey: rate your knowledge, get a dynamic curriculum, not a six-month-old PDF syllabus.
Hasbe lights up—Walmart uses Neo4j for employee knowledge graphs across two million workers; Quarles & Brady converts unstructured legal corpora into graphs for problem-specific paths. The pattern is the same: unstructured source material becomes interconnected concepts; the system routes learners (or agents) along the shortest meaningful path to competence.
Why Vectors Are Not Enough
Wilson’s three-little-pigs example lands cleanly: vector search cannot infer that a story about pigs is a children’s tale when the words kid and story never appear. Ideas live in relationships, not token neighborhoods. Hasbe agrees—similarity blobs miss interconnectedness; graph paths expose what you know, what you do not, and what to learn next.
The episode closes on self-improving systems: low-fidelity models that admit ignorance, ask humans, and write the answer back into the graph; voice-first platforms using native graph storage as a latency cache; virtual versus native graphs chosen by SLA, not religion. Knowing before doing is not a slogan—it is the product requirement for the next generation of software Hasbe sees customers already building.
Open the Ideation hub, or read the full episode transcript.