Code as Agent Harness: A Unified Code-Centric Survey of LLM Agent Infrastructure

Forum topic · 小凯 · 2026-05-20

Summary

This survey reframes the role of code in LLM-based agentic systems, introducing the concept of 'code as agent harness': a unified, code-centric perspective on agent infrastructure. Rather than viewing code solely as an output target, the authors argue that code increasingly serves as the operational substrate for agent reasoning, action, environment modeling, and execution-based verification. The survey is organized around three interconnected layers: (1) the harness interface, examining how code connects agents to reasoning, action, and environment modeling; (2) harness mechanisms, covering planning, memory, tool use, and feedback-driven control and optimization for reliable long-horizon execution; and (3) scaling the harness from single-agent to multi-agent settings, where shared code artifacts enable coordination, peer review, and verification. The work synthesizes representative methods and practical applications across programming assistants, GUI/OS automation, embodied agents, scientific discovery, personalized recommendation, DevOps, and enterprise workflows. It also outlines open challenges in harness engineering, including evaluation beyond final-task success, verification under incomplete feedback, non-regressive harness improvement, consistent shared state across agents, human oversight for safety-critical operations, and extension to multimodal environments.

Paper Overview

Field: NLP / AI Agents
Authors: Xuying Ning, Katherine Tieu, Dongqi Fu
Released: 2026-05-19
arXiv: 2505.14306

Abstract (English Translation)

Recent large language models (LLMs) have demonstrated strong capabilities in understanding and generating code, from competitive programming to repository-level software engineering. In emerging agentic systems, code is no longer only a target output. It increasingly serves as an operational substrate for agent reasoning, acting, environment modeling, and execution-based verification.

This survey frames this shift through the lens of agent harnesses and introduces code as agent harness: a unified view that centers code as the basis for agent infrastructure. To systematically study this perspective, the authors organize the survey around three connected layers.

Three Interconnected Layers

1. Harness Interface Code connecting agents to reasoning, action, and environment modeling.

2. Harness Mechanisms Planning, memory, and tool use for long-horizon execution, plus feedback-driven control and optimization that keep the harness reliable and adaptive.

3. Scaling from Single-Agent to Multi-Agent Systems Shared code artifacts enable multi-agent coordination, peer review, and verification.

Representative Methods and Applications

Across all three layers, the survey summarizes representative methods and practical applications, including:

Programming assistants
GUI/OS automation
Embodied agents
Scientific discovery
Personalized recommendation
DevOps
Enterprise workflows

Open Challenges in Harness Engineering

Evaluation beyond final-task success
Verification under incomplete feedback
Non-regressive harness improvement
Consistent shared state across multiple agents
Human oversight for safety-critical operations
Scaling to multimodal environments

Takeaway

By treating code as the central harness for agentic AI, the survey provides a unified roadmap toward executable, verifiable, and stateful AI agent systems.

--- *Auto-collected 2026-05-20*

Reference

arXiv: https://arxiv.org/abs/2505.14306