- 2026/05/24: Released the paper and codebase. The second version of the paper has also been submitted to arXiv, and the code release includes the evolution prompts used to build the harness.
Life-Harness is the code release for "Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents." It targets a practical question: when a frozen LLM agent repeatedly fails in a deterministic environment, can we improve the runtime harness around the agent instead of retraining the model or modifying the environment?
The answer is yes. Life-Harness turns recurring failures into reusable runtime interventions across action realization, environment contracts, trajectory regulation, and procedural skills. The model remains frozen; the benchmark environment remains intact; only the harness interface adapts.
| Benchmarks | Model backbones | Settings improved | Avg. relative gain | Training-free |
|---|---|---|---|---|
| 7 | 18 | 116 / 126 | 88.5% | Yes |
| What changes? | What stays fixed? | Why it matters |
|---|---|---|
| Runtime harness behavior | LLM weights | No finetuning or model-specific training pipeline |
| Prompted environment interface | Benchmark environment | Keeps deterministic evaluation comparable |
Across 7 deterministic agent benchmarks and 18 model backbones, Life-Harness improves 116 / 126 model-environment settings, with an 88.5% average relative improvement reported in the paper.
Life-Harness evolves a small set of runtime layers from observed failures, then reuses those layers during evaluation.
| Harness flag | Paper layer | Runtime role |
|---|---|---|
h2 |
Action Realization Layer | Helps convert model decisions into executable environment actions. |
h3 |
Environment Contract Layer | Makes task and environment constraints explicit at runtime. |
h4 |
Trajectory Regulation Layer | Regulates multi-step interaction traces to avoid repeated failure patterns. |
h5 |
Procedural Skill Layer | Reuses procedural knowledge distilled from recurring successful recoveries. |
When the harness is disabled, these layers are not applied.
This repository keeps the two benchmark families in separate folders because their environments and dependencies are intentionally different.
| Suite | Environments | Start here |
|---|---|---|
| AgentBench-style harness | ALFWorld, DBBench, OS, WebShop | AgentBench/README.md |
| tau-bench-style harness | Airline, Retail, Telecom | TauBench/README.md |
Life-harness/
AgentBench/ # Docker-based AgentBench-style tasks
TauBench/ # uv-based tau-bench-style tasks
assets/ # README figures
Clone the repository, then enter the benchmark suite you want to run:
cd Life-harness
# tau-bench-style tasks: Airline, Retail, Telecom
cd TauBench
# AgentBench-style tasks: ALFWorld, DBBench, OS, WebShop
cd ../AgentBenchEach subfolder README contains its own environment setup, evaluation commands,
and harness switches. API keys and provider URLs should be configured locally
through environment variables or .env files; do not commit them.
If you use this repository, please cite the paper:
@article{xu2026adapting,
title={Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents},
author={Xu, Tianshi and Wen, Huifeng and Li, Meng},
journal={arXiv preprint arXiv:2605.22166},
year={2026}
}