Life-Harness: Runtime Adaptation for Deterministic LLM Agents
Life-Harness is an innovative technique that enhances frozen LLM agents without altering model weights or evaluation contexts. This method modifies the runtime harness, which serves as the interface for observation, tool utilization, action execution, feedback analysis, and trajectory management, by transforming frequent interaction failures into reusable solutions. In trials conducted across seven deterministic environments from τ-bench, τ²-bench, and AgentBench, Life-Harness showed improvements in 116 out of 126 model-environment combinations involving 18 models. This strategy tackles issues in rule-based domains stemming from discrepancies at the model-environment interface, rather than adjusting model parameters.
Key facts
- Life-Harness is a lifecycle-aware runtime harness for frozen LLM agents.
- It does not change model weights or evaluation environments.
- It evolves from training trajectories by converting recurring failures into interventions.
- Interventions span environment contracts, procedural skills, action realization, and trajectory regulation.
- Tested on seven deterministic environments from τ-bench, τ²-bench, and AgentBench.
- Improved 116 out of 126 model–environment settings across 18 models.
- Focuses on model–environment interface mismatches in rule-governed domains.
- The method remains fixed during held-out evaluation.
Entities
—