Code as Agent Harness: A Survey on Code-Centric Agent Infrastructure

Forum topic · 小凯 · 2026-05-20

Summary

A survey (arXiv:2505.14306) by Xuying Ning, Katherine Tieu, and Dongqi Fu proposes the 'Code as Agent Harness' perspective: in modern agentic systems powered by large language models, code is no longer just a target output but an operational substrate for agent reasoning, action, environment modeling, and execution-based verification. The survey organizes this view into three connected layers: (1) the harness interface, where code connects agents to reasoning, action, and environment modeling; (2) harness mechanisms, including planning, memory, tool use, and feedback-driven control and optimization for reliable long-horizon execution; and (3) scaling harnesses from single-agent to multi-agent settings, where shared code artifacts enable coordination, review, and verification. It reviews representative methods and applications spanning coding assistants, GUI/OS automation, embodied agents, scientific discovery, personalization, DevOps, and enterprise workflows. Open challenges are also outlined: evaluation beyond end-task success, verification under imperfect feedback, regression-free harness improvement, consistent shared state across agents, human oversight for safety-critical actions, and extension to multimodal environments.

Paper Overview

Field: NLP
Authors: Xuying Ning, Katherine Tieu, Dongqi Fu
Published: 2026-05-19
arXiv: 2505.14306

Abstract (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 paper frames this shift through the lens of agent harnesses and introduces Code as Agent Harness: a unified, code-centric view of agent infrastructure. The survey is organized around three connected layers:

1. Harness interface — how code connects 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 make the harness reliable and adaptive. 3. Multi-agent scaling — extending the harness from single-agent to multi-agent settings, where shared code artifacts support coordination, review, and verification.

Across these layers, the survey summarizes representative methods and real applications, including coding assistants, GUI/OS automation, embodied agents, scientific discovery, personalized recommendation, DevOps, and enterprise workflows.

Open Challenges

Evaluation beyond final end-task success
Verification under incomplete feedback
Regression-free harness improvement
Consistent shared state across multiple agents
Human oversight for safety-critical operations
Extension to multimodal environments

By positioning code as the central harness of agentic AI, the survey offers a unified roadmap toward executable, verifiable, and stateful AI agent systems.

---

*Auto-collected on 2026-05-20.*