Growing Harness: Turning Repeated LLM Agent Control into Reusable Executable Code

Paper Overview

• Field: Agents
• Authors: Laizhen Li, Jiarui Li, Juanjuan Zhao, Kejiang Ye, Ye Li et al.
• Published: 2026-09-22
• arXiv: 2609.26760

Motivation

LLM agents often handle streams of related tasks, yet standard harnesses repeatedly ask the model to reconstruct the same control decisions inside each task's context. The authors study whether task feedback can turn recurring control into reusable executable code, reserving LLM calls for task-specific semantic reasoning.

Method: Growing Harness

Growing Harness is a failure-guided training paradigm that learns the agent harness itself from a strategy-free scaffold — one that exposes fixed model and tool interfaces but encodes no task-solving controller. Key components:

1. Function-level execution traces localize each failure to a bounded code surface. 2. Joint repair: an optimizer repairs a window of failures together. 3. Success-first held-out gate: rolls back repair sequences that harm prior capability. 4. Cumulative growth: accepted edits accumulate in one shared harness, letting its control structure emerge from task feedback.

Results

Evaluated on BrowseComp-Plus and WebArena-Verified with three deployment models from 4B to 120B parameters:
• Highest mean success rate in five of six benchmark-model settings; trails the best mean by only 0.7 pp. in the sixth.
• Relative to a Tool-Calling agent: 76.0-91.8% fewer LLM calls and 74.4-98.6% lower deployed-agent inference cost.
• On WebArena-Verified, success remains 44.7-45.3% across model scales, whereas Tool-Calling falls to 6.7% with the 4B model.
• Ablations show trace-local edits, joint repair, and gate-based rollback each improve final success.

Takeaway

Persistent program growth can move recurring control out of model context and into low-cost code, yielding reusable specialist agents that remain effective even with smaller deployment models.

Original Abstract

> Large language model (LLM) agents often handle streams of related tasks, yet standard harnesses repeatedly ask the model to reconstruct the same control decisions inside each task's context. We study whether task feedback can instead turn recurring control into reusable executable code, while reserving LLM calls for task-specific semantic reasoning. We introduce Growing Harness, a failure-guided training paradigm that learns the agent harness itself from a strategy-free scaffold that exposes fixed model and tool interfaces but encodes no task-solving controller. Function-level execution traces localize each failure to a bounded code surface, an optimizer repairs a window of failures jointly, and a success-first held-out gate rolls back repair sequences that harm prior capability. Accepted edits accumulate in one shared harness, allowing its control structure to emerge from task feedback. Across BrowseComp-Plus and WebArena-Verified with three deployment models from 4B to 120B parameters, Growing Harness achieves the highest mean success in five of six benchmark-model settings and trails the best mean by 0.7 pp. in the sixth. Relative to a Tool-Calling agent, it reduces LLM calls by 76.0-91.8% and deployed-agent inference cost by 74.4-98.6%. On WebArena-Verified, its success remains 44.7-45.3% across model scales, whereas Tool-Calling falls to 6.7% with the 4B model. Ablations show that trace-local edits, joint repair, and gate-based rollback each improve final success. These results show that persistent program growth can move recurring control out of model context and into low-cost code, yielding reusable specialist agents that remain effective with smaller deployment models.

--- *Auto-collected on 2026-09-24*