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

Paper · arXiv 2608.21156 · Published August 21, 2026
LLM Agents

Abstract Large language models (LLMs) have rapidly evolved from language generation models into autonomous agents capable of solving increasingly complex and long-horizon tasks. This evolution has been accompanied by a series of emerging engineering paradigms, including Prompt Engineering for eliciting model capabilities, Context Engineering for managing information access, Harness Engineering for organizing external tools and resources, and Loop Engineering for enabling continual reflection and self-improvement. However, as real-world tasks grow in complexity, a fundamental limitation of individual intelligence emerges: many tasks inherently require heterogeneous expertise, interdependent subtasks, parallel execution, independent verification, and persistent state, and these requirements exceed the organizational capacity of any single agent. Simply augmenting an individual agent’s capabilities or context cannot resolve this architectural mismatch. Instead, intelligence must be distributed across multiple specialized agents and organized at the system level. We refer to this capability as System Intelligence: the ability of an agent system to organize and coordinate multiple intelligent components into a coherent, adaptive whole that pursues a shared objective.

Introduction. Large language models (LLMs) have rapidly evolved into a foundational component of modern intelligent systems, driven by substantial advances in language understanding, reasoning, generation, and decision making [31, 54, 63, 130, 350, 352]. This progress has largely followed two complementary directions: strengthening the capabilities encoded in model parameters during training [58, 235, 281] and improving how these capabilities are activated and utilized at inference time [11, 28, 304, 387]. Specifically, early research primarily focused on the former, using large-scale pre-training and post-training to expand and refine the knowledge and reasoning capabilities of individual models [31, 63, 235, 386]. More recently, increasing attention has shifted toward inference-time engineering, where Prompt Engineering and Context Engineering serve as complementary approaches for shaping model behavior.

Discussion / Conclusion. Large language models have rapidly evolved from standalone generators into individual agents capable of sustained interaction, tool use, and iterative execution. Yet, as tasks become more heterogeneous, interdependent, and longhorizon, the limitations of individual intelligence become increasingly clear: a single agent loop struggles to support parallel work, specialized expertise, independent verification, and persistent state. This survey argues that the next frontier is System Intelligence, the ability of an agent system to organize complex objectives, coordinate heterogeneous components, and maintain coherent runtime state across the task lifecycle. To support this transition, we introduce Graph Engineering as a structure-centered engineering paradigm that uses graph abstractions to make system relations explicit, operational, and adaptable. We organize the literature around three complementary graph views: work organization, agent coordination, and runtime state management.

Synthesis notes that discuss concepts related to this paper 7

Notes in this collection most semantically related to the paper, ranked by cosine similarity.

Lines of inquiry this paper opens 24

Research framings built by reading the notes related to this paper — the questions it feeds into.

How do evaluation biases undermine LLM quality assessment systems? How can LLM user simulators model realistic goal-driven conversation? How should we design LLM systems to maintain alignment and control? Does self-reflection enable models to reliably correct their errors? What articulatory information do speech signals carry that text cannot? How do knowledge injection methods compare across cost and effectiveness? How faithfully do LLMs reflect their actual reasoning in outputs and explanations? Why do language models reinforce false assumptions instead of correcting them? Do accurate-looking LLM outputs hide structural failures in learning and reasoning? How can models identify insufficient information and respond appropriately without guessing? How do language models inherit human biases from training data? What capability tradeoffs emerge when scaling model reasoning abilities? How should inference compute be adaptively allocated based on prompt difficulty? What pretraining choices and baseline capability constrain reinforcement learning gains? Why do continual learning scenarios trigger catastrophic forgetting and interference? Why does finetuning cause catastrophic forgetting of model capabilities? How do training priors constrain what context information can override? How do language models establish social grounding in human dialogue? Can AI-generated outputs constitute genuine knowledge or valid claims? How do multi-agent systems achieve genuine cooperation and reasoning? What drives capability and cost efficiency in agent systems? Can prompting strategies overcome LLM biases without model fine-tuning?