Graph Engineering in the Era of LLM Agents: From Individual Intelligence to System Intelligence
Abstract Overview
This paper provides a comprehensive survey on how LLM-based systems are transitioning from single-agent individual intelligence to multi-component system intelligence to tackle complex, long-horizon tasks. The authors identify key limitations of single-agent loops—including difficulties in handling interdependent parallel subtasks, specialized verification, and persistent execution state—and introduce Graph Engineering as a system-level organizational paradigm. Graph Engineering employs explicit, dynamic graph representations across three core dimensions: Task Organization, Agent Coordination, and Runtime State Management. The survey traces foundational engineering stages from prompt and context engineering to harness and loop engineering, while categorizing emerging benchmarks, open-source libraries, practical applications, and future research directions such as ontology engineering and graph-native agent operating systems.
Novelty
The paper formalizes Graph Engineering as a distinct system-level engineering paradigm for LLM agent architectures, structuring multi-agent coordination around the transition from individual to system intelligence. It contributes a unifying taxonomy defined by three interconnected graph pillars: Task Organization, Agent Coordination, and Runtime State Management.
Results
As a survey and conceptual framework, the paper's primary achievement is a systematic synthesis of the agent engineering ecosystem rather than a standalone empirical model. It unifies existing benchmarks, datasets, open-source libraries, and domain applications under a coherent taxonomy while detailing technical challenges in failure localization, state recovery, and cross-run system evolution.
Key Points
- The paper delineates the progression from Model Intelligence and Individual Intelligence to System Intelligence, demonstrating why complex tasks require explicit system-level organization beyond single-agent loops.
- It establishes the Graph Engineering paradigm around three operational functions: structuring tasks and workflows, coordinating heterogeneous agent teams, and managing runtime state for monitoring, fault localization, and recovery.
- The survey systematically catalogues benchmarks, software libraries, and domain applications, while outlining future directions including self-evolving graph systems, graph-native agent operating systems, and ontology engineering.