Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents

Researchers introduce Life-Harness, a runtime interface adaptation method that improves frozen LLM agent performance without modifying model weights. The technique evolves from training trajectories to fix model-environment mismatches, achieving 88.5% average improvement across 126 settings and demonstrating cross-model transferability that suggests environment-side structure matters as much as model architecture.

Life-Harness represents a paradigm shift in how researchers approach LLM agent optimization. Rather than continuously fine-tuning model parameters—the dominant approach in agent adaptation—this work identifies that many agent failures stem from interface mismatches between language models and their execution environments. By keeping models frozen and instead adapting the runtime harness layer, the method decouples agent improvement from computationally expensive model retraining.

The research addresses a practical bottleneck in deterministic domains like code execution, planning, and structured reasoning. These environments have strict rules and expected behaviors that models often misinterpret or fail to execute correctly. Life-Harness converts recurring interaction failures into reusable interventions at four integration points: environment contracts, procedural skills, action realization, and trajectory regulation. This structured approach to interface design reflects growing recognition that LLM agents function as socio-technical systems where the glue matters as much as the components.

The transferability findings carry significant implications for model deployment and economics. Harnesses evolved from a single 4B parameter model (Qwen3-4B-Instruct) transfer effectively to 17 other models, suggesting that environment-specific patterns are learnable independently of model size or architecture. This reduces the need for model-specific optimization pipelines and makes agent development more generalizable across diverse model choices.

For practitioners building production agents, Life-Harness opens a path to performance improvements without incurring inference costs or retraining overhead. The approach also democratizes agent optimization—smaller organizations can leverage harness adaptation rather than requiring resources for large-scale model fine-tuning. Looking ahead, this work may catalyze broader investigation into interface-first agent design, potentially reshaping how teams architect reliable LLM systems.

• →Life-Harness improves frozen LLM agents by 88.5% on average by adapting runtime interfaces rather than model weights
• →The method converts recurring interaction failures into reusable interventions across environment contracts, procedural skills, action execution, and trajectory control
• →Harnesses evolved from a single 4B model transfer effectively to 17 other models, indicating environment-side structure generalizes across architectures
• →Runtime interface adaptation provides a computationally efficient alternative to parameter-based model fine-tuning for agent optimization
• →Results span seven deterministic benchmark environments with 126 model-environment configurations, demonstrating broad applicability