Graph engineering
What it is
Graph engineering is the practice of designing a multi-agent AI system as an explicit graph rather than a single freeform agent loop. Nodes represent agents, tools, deterministic functions, routers, and human checkpoints; edges represent the transitions that are actually permitted between them. Frameworks like LangGraph operationalize this by treating agent execution as graph traversal with typed state, checkpoints, and controlled cycles, so persistence, recovery points, and human-approval gates are placed deliberately in the topology instead of being left to an agent's own in-the-moment reasoning.
Why it's emerging now
"Loop engineering" had already established that production agent systems need a repeatable, engineered structure rather than one-off prompting — but a single loop only describes how one agentic node executes, not how several agents, tools, and human checkpoints should hand off work reliably as systems scale to multi-agent. The term caught on fast in mid-to-late July 2026, reportedly sparked by developer Peter Steinberger publicly asking whether the field had already moved from talking about loops to talking about graphs, and was picked up independently across practitioner blogs through August. By early September it had been folded into GitHub's own engineering blog as one of the new terms developers are expected to know, alongside harness engineering and loop engineering.
How it relates to what you know
It sits one level above loop engineering in your ledger: loop engineering designs how a single agentic node executes repeatedly, while graph engineering designs the topology that connects multiple such loops, agents, tools, and humans together — the next step in the pattern of prompt engineering → context engineering → loop engineering, where the engineering leverage keeps moving further out from the model itself. It's also worth distinguishing from knowledge graphs/GraphRAG, also already on your ledger: those model what a system knows for retrieval, while graph engineering models how a system runs.