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

· Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Life-Harness is a novel lifecycle-aware runtime harness designed to enhance the performance of frozen Large Language Model (LLM) agents in deterministic, rule-governed environments. This system adapts the model–environment interface rather than modifying LLM weights or evaluation environments. It evolves by converting recurring interaction failures from training trajectories into reusable interventions across four layers: environment contracts, procedural skills, action realization, and trajectory regulation. Evaluated on seven deterministic environments from τ-bench, τ²-bench, and AgentBench, Life-Harness improved 116 out of 126 model–environment settings across 18 model backbones, achieving an average relative improvement of 88.5%. Harnesses evolved using Qwen3-4B-Instruct successfully transferred to 17 other models, demonstrating broad applicability.

Key takeaway

For Machine Learning Engineers developing LLM agents for deterministic, rule-governed environments, you should prioritize runtime interface adaptation as a powerful complement to model-centric training. Implementing lifecycle-aware harness layers, such as those for environment contracts, action realization, or trajectory regulation, can significantly improve agent performance and generalization across various model backbones without modifying model weights. This approach offers substantial gains, even enabling smaller models to compete with larger, specialized ones.

Key insights

Adapting the LLM agent's runtime interface, not its model weights, significantly boosts performance in deterministic environments.

Principles

Method

Life-Harness evolves by diagnosing recurring interaction failures in training trajectories and converting them into fixed, reusable interventions across four lifecycle layers: Environment Contract, Procedural Skill, Action Realization, and Trajectory Regulation, assisted by a coding agent.

In practice

Topics

Code references

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.