Source note
Code as Agent Harness
Software Foundation ModelsCode IntelligenceAgentic AIMulti Agent Software EngineeringHuman AI Interaction
Summary
This survey argues that code is becoming the harness for AI agents: the layer where agents reason, act, store state, call tools, and verify work. It matters for software engineering, OS automation, DevOps, and multi-agent workflows because these systems need executable state and checks, not only generated text.
Problem
- LLM agents often need to act over long tasks, use tools, remember context, and check their own work; plain chat outputs do not give enough control or traceability.
- Agent systems need safer ways to connect reasoning with actions in software repositories, GUIs, operating systems, enterprise workflows, and scientific work.
- Multi-agent systems need shared artifacts and consistent state so agents can coordinate, review, and verify each other's work.
Approach
- The paper is a survey, not a new model or benchmark paper.
- Its core mechanism is simple: treat code as the agent harness, meaning code holds the agent's state, tool calls, plans, environment models, and execution checks.
- It organizes the topic into 3 layers: harness interfaces, harness mechanisms, and scaling from single-agent to multi-agent systems.
- It reviews methods for planning, memory, tool use, feedback-driven control, optimization, coordination, review, and verification.
- It maps the idea across coding assistants, GUI and OS automation, embodied agents, scientific discovery, personalization, DevOps, and enterprise workflows.
Results
- The excerpt reports no benchmark scores, datasets, ablation studies, or measured performance gains.
- The main concrete claim is a 3-layer organization of code-based agent harnesses: interface, mechanisms, and multi-agent scaling.
- It identifies 6 open engineering problems: evaluation beyond final task success, verification with incomplete feedback, regression-free harness improvement, consistent shared state, human oversight for safety-critical actions, and multimodal extension.
- It covers 7 application areas named in the abstract: coding assistants, GUI/OS automation, embodied agents, scientific discovery, personalization and recommendation, DevOps, and enterprise workflows.
- Its claimed contribution is a survey-level synthesis for executable, verifiable, and stateful agent systems rather than a quantitative breakthrough.