Beyond Model Capability: The Harness for SE Agents

The promise of foundation models in automated code generation has outpaced their practical application in realistic software engineering settings. Autonomous agents, while powerful, remain unreliable. This paper challenges the prevailing narrative that limitations lie solely within the foundation model itself.

Visual TL;DR. Agent Unreliability problem Beyond Model Capability. Beyond Model Capability solution AI Harness System. AI Harness System enables Systemic Capability. AI Harness System details Harness Responsibilities. AI Harness System enables Verifiably Correct Changes. Verifiably Correct Changes leads to Redefined Success.

Agent Unreliability: autonomous software engineering agents are currently unreliable in practice
Beyond Model Capability: limitations are not solely within the foundation model itself
AI Harness System: novel intermediary system for agents to perceive, act, and get feedback
Systemic Capability: capability emerges from model, harness, and development environment interplay
Harness Responsibilities: eleven key responsibilities including task spec, context, tools, memory, verification
Verifiably Correct Changes: enables autonomous agents to make verifiably correct software changes
Redefined Success: redefining success in autonomous software engineering beyond model prowess

Visual TL;DRQuickExplainDeeper

!-- /sh-diagram -->

The Systemic Nature of Software Engineering Capability

The researchers propose that effective software engineering capability emerges from the interplay between a foundation model, a mediating harness, and the development environment. This AI Harness acts as a critical intermediary, dictating how an agent perceives a project, executes actions, receives feedback, and confirms task completion. This reframes the problem from individual model prowess to the architecture of the entire system. The harness is formalized with eleven key responsibilities, including task specification, context selection, tool access, project memory, and verification.

A Ladder of Runtime Support for Autonomous Agents

To operationalize this concept, the paper introduces a four-level harness ladder (H0-H3). Each level incrementally exposes more runtime support to the agent. This graduated approach allows for systematic evaluation and development. The framework's evaluation protocol generates auditable 'episode packages,' which vary in their evidence structure based on the harness level. Higher levels yield richer outputs, such as reproduction logs, failure attributions, and structured verification reports, moving beyond simple patch generation for foundation model software engineering.

Redefining Success in Autonomous Software Engineering

The core thesis shifts the central question from 'can a foundation model produce a patch?' to 'can the model-harness-environment system produce a verifiably correct, attributed, and maintainable change?' This systemic view is crucial for advancing the field of foundation model software engineering. The paper concludes by outlining a research agenda focused on the necessary runtime systems for future autonomous software agents.