AIAny
Icon for item

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

Proposes “Graph Engineering”: using explicit, dynamic graphs to represent tasks, agents, tools, and system state so LLM-based agent systems can coordinate, persist, and evolve. Surveys principles, methods, applications, and curates related resources.

Introduction

Most work on LLM agents optimizes prompt design, context windows, tool wrappers, or reflection loops for a single agent. This paper argues that those individual-level fixes hit an architectural ceiling for complex, heterogeneous, long-horizon tasks — what’s needed instead is system-level structure: explicit, evolving graphs that encode tasks, agents, tools, and state as first-class objects.

Key Findings
  • Conceptual framing: introduces "Graph Engineering" and the notion of System Intelligence — the ability of a multi-component agent system to organize, coordinate, and adapt toward shared objectives using graph-shaped substrates. This reframes coordination, memory, planning, and tool orchestration as graph design problems rather than purely prompt or model problems.
  • Roles for graphs: identifies recurring graph abstractions (task/workflow graphs, tool graphs, knowledge/knowledge-augmented graphs, agent-communication/coordination graphs, and state/memory graphs) and explains how each addresses specific agent limitations such as fragmentation of context, unreliable long-term memory, and brittle tool management.
  • Methodological guidance: surveys approaches for integrating graphs with LLM-based agents — from prompting and retrieval-augmented workflows to light-weight graph neural modules, schema-constrained graph rewrites, and human-in-the-loop graph editing — and highlights dynamic-graph learning as infrastructure for continuous agent evolution.
  • Practical trade-offs and engineering patterns: argues that graphs improve reliability, interpretability, efficiency, and modularity, but introduce schema design, update semantics, consistency maintenance, and evaluation challenges. The paper collects an accompanying resource set (curated repo of papers, datasets, and projects) to lower adoption friction.
Who It's For and Trade-offs

Great fit if you design multi-agent workflows, need persistent and verifiable execution state across long-horizon tasks, or must coordinate specialized agents/tools in parallel and with independent verification. The paper is most useful for researchers and engineering teams building orchestrated LLM-agent systems, MAS, or workflow engines that require explicit state, provenance, and structured coordination.

Look elsewhere if your use case is single-turn prompt engineering, small-scale tool wrappers, or one-off assistant interactions — the graph engineering stack brings nontrivial design and maintenance cost and is overkill for ephemeral or trivially decomposable tasks. Expect upfront effort on schema definition, consistency policies, and runtime graph maintenance.

Practical Takeaway

If your system needs dependable coordination, transparent decision paths, and evolving capabilities across heterogeneous agents, treat graphs as primary design artifacts: choose clear node/edge types, define schema-driven update rules, instrument graph change traces for debugging, and combine lightweight graph models with LLM prompts rather than attempting to absorb all structure into prompts alone.

Information

  • Websitearxiv.org
  • AuthorsYuyuan Feng, Zhishang Xiang, Chaobin Yang, Qichao Ma, Zerui Chen, Yujing Zhang, Ke Huang, Chuanjie Wu, Zhaoxu Liu, Yili Wang
  • Published date2026/08/21

Categories

More Items

Evaluates whether coding agents can modify real scientific software while preserving domain-specific scientific contracts. Contains 119 repository-level tasks across 98 GitHub projects and 20 scientific domains, measures reproducible edits in pinned Docker images, and analyzes recurring failure modes.

Proposes FACET, a framework that synthesizes verifiable terminal tasks by reconstructing scenario intent and grounding instruction, solution, and verifier in a shared executable container state. Key features include environment-first generation, execution-based validation, and targeted repair to preserve source intent and cross-artifact consistency.

Wraps static, hand-built environments with a programmable plug-in harness that reshapes environment behavior without changing underlying logic. EnvRigger automates diagnosis and synthesis of harness components from agent failure trajectories, validating edits via fresh rollouts to improve agent success and efficiency.