Rakesh Gohel, who is building a governance infrastructure for AI agents at JUTEQ, argues that every decision an agent makes rests on three questions: what is true, what matters now, what to do next. A knowledge graph answers the first well; each of the other two, in his view, deserves its own graph.
He takes the example of a simple request made to an airline agent: rebook a passenger. The knowledge graph holds what is true at the company-wide scale: passengers, aircraft, flights, fares, routes, airports, weather, policies, kept stable and governed. The context graph extracts from it a small slice limited to the decision at hand: Gate C8, Flight 34/185, four seats left, nine minutes before closing, and how similar cases have been handled. The infographic sums up this principle: don't retrieve everything, scope the graph to what the decision needs.
The procedural graph stores procedures as procedure-relation-procedure triples. The agent moves from checking availability to comparing alternatives then to resolution, checks the visa if needed, applies fare rules when options exist, and escalates to a human when there are none. The graph guides the next step without dictating it: the agent reasons over the scoped context, follows the next edge, and remains free to deviate from it if the graph is wrong.
Every trace gets written back. Precedents refine the context, failed runs correct the procedure, and recurring patterns become governed facts. The infographic embodies this loop through a "Trace & Learn" store that collects success and failure traces and updates knowledge, context, and procedure.
It adds an element absent from the text, the "Semantic Spine," a five-link chain: the ontology says what things are, the knowledge graph what's true, the context graph what matters now, the procedural graph what to do next, and the agent acts and writes the result back.
The author's lesson comes down to three instructions: scope the graph at each step, store the procedure as a graph rather than as a prompt, and log every decision, because that log becomes the audit trail. The post closes by asking the reader which of the three graphs their agent is missing.