Life Harness: a runtime interface for rule-governed AI agents
A base language model does not come with durable state, tool permissions, or direct access to an application’s rules. A runtime supplies those things. It can give the model observations, accept or reject actions, call tools, and return feedback.
Life Harness is a research framework for this runtime layer. The paper evaluates frozen LLM agents in seven deterministic benchmark environments, including tau-bench, tau²-bench, and AgentBench. It is not a general design for physical-world robotics, and its results should be read in the scope of those environments.
The four components
The paper describes four parts of the harness. They do not form one simple filter. Each intervenes at a different point in the agent loop.
The environment contract states task-specific constraints and available actions before an agent proceeds. Procedural skill captures useful procedures from training trajectories. Action realization turns a proposed action into the format the environment expects. Trajectory regulation looks at the action history and can intervene when a pattern appears unproductive.
Together, these components provide more structure around a model call. They can reduce interface-driven mistakes in an environment with clear rules, but they do not guarantee a correct plan, safe outcome, or successful tool call.
What “without retraining” means here
The base model’s weights remain frozen in the study. The benchmark environments remain unchanged too. The harness itself is evolved from training trajectories, then held fixed for evaluation. That is different from updating the model weights, but it is not the same as a system that learns a new procedure online after every successful task.
This distinction is useful in product work. A team can often improve an agent by changing its prompts, action schemas, tool wrappers, checks, and stored procedures without training a new foundation model. Whether those changes help still depends on the task and on evaluation against failures that matter.
Results and their limits
The authors report improvements in 116 of 126 model-environment settings across 18 backbones and seven deterministic environments. They report an average relative improvement of 88.5%. They also report transfer from a harness evolved on Qwen3-4B-Instruct trajectories to 17 other tested backbones.
Those are promising research results, not a general production guarantee. The reported transfer concerns the tested models and environments. Other tools, stateful systems, safety requirements, and real-world side effects need their own controls and evaluation.
Why the interface still matters
It is useful to separate language understanding from the mechanics of acting in an application. The model can interpret a request and propose a plan. The surrounding system should define tool schemas, permissions, validation, error handling, logs, and human approval where appropriate.
Life Harness offers one way to think about that boundary. It does not replace a model with a set of rules. It gives the model a more structured way to operate in a particular environment, while keeping the surrounding runtime visible enough to test and improve.
Read the Life Harness paper and its official code for the experimental setup and implementation.