Graph Engineering in the Era of LLM Agents: From Individual Intelligence to System Intelligence

The central argument here is architectural: single-agent loops are structurally inadequate for tasks that require parallel execution, heterogeneous expertise, independent verification, and persistent state. No amount of prompt tuning or context expansion fixes this mismatch. The solution the paper proposes is Graph Engineering — using explicit graph structures to organize tasks, coordinate agents, and manage runtime state across a distributed system.

The paper traces a clean progression. Model Intelligence lives inside a single inference call. Individual Intelligence extends that call into a persistent loop with tools, memory, and skills. System Intelligence is the next level: multiple specialized agents organized as a coherent whole, with explicit structures governing who does what, in what order, and with what shared state. The authors argue that most current multi-agent systems fail to reach this level not because they lack agents, but because they lack the organizational substrate — clear task dependencies, capability graphs, communication topologies, and traceable runtime state.

The three-part framework is the paper's real contribution. Task Organization covers how a high-level objective becomes a schedulable graph of subtasks with explicit dependencies. Agent Coordination covers capability modeling, team topology, and communication patterns — distinguishing stable role assignments from the dynamic information flows that emerge during execution. Runtime State Management covers state recording, fault localization, and failure recovery, arguing that distributed execution requires something closer to transactional semantics than a shared conversation history.

A fourth dimension, System Evolution, addresses how execution evidence should feed back into persistent structural changes — not just rerouting within a fixed workflow, but actually revising task graphs, agent assignments, and capability relations across runs. The paper is candid that this remains largely aspirational: most production frameworks support dynamic routing within predefined structures, but cross-run structural self-modification is rare.

The survey also introduces Ontology Engineering as a semantic layer above Graph Engineering — shared, machine-interpretable definitions of goals, roles, states, and evidence. Without it, graph structures can be explicit without being consistently interpreted across components or executions.

The coverage is genuinely broad, spanning benchmarks, open-source libraries, and applications from software engineering to clinical decision support to economic simulation. The application survey's honest conclusion is worth noting: Work Organization and Agent Coordination are increasingly common in practice, but persistent System Evolution — where runtime evidence actually changes future organizational structure — remains rare across all domains surveyed.