Life-Harness Runtime Adaptation

• Life-Harness is a lifecycle-aware runtime harness designed to adapt environment interfaces in deterministic LLM agents, addressing persistent mismatch failures.
• It evolves a four-layered runtime scaffold to mediate observation structuring, tool invocation, action validation, and trajectory regulation.
• Empirical results show up to 120% relative improvement, demonstrating robust, model-agnostic performance across diverse deterministic environments.

Life-Harness is a lifecycle-aware runtime harness designed to address persistent model–environment interface failures in deterministic, rule-governed agent domains. Rather than modifying LLM parameters, Life-Harness adapts the runtime scaffold—mediating observation structuring, tool invocation, action validation, feedback interpretation, and trajectory regulation—by evolving reusable, environment-specific interventions. This approach yields substantial, cross-model improvements in frozen LLM agents for environments including household simulators, web shopping, operating system control, and text-to-SQL tasks (Xu et al., 21 May 2026).

1. Problem Formulation and Motivation

In deterministic agent environments (e.g., τ-bench, τ²-bench, AgentBench), LLM agent failures often stem from model–environment interface mismatches, independent of the model’s latent reasoning capacity. These mismatches comprise disorganized or incomplete observations (e.g., missing admissible-action lists), misinterpretation of tool schemas (argument names, types, ordering), inability to canonicalize output into executable formats (such as JSON parse errors or missing arguments), and failures to convert feedback (error messages, "nothing happens") into actionable recovery signals. Traditional model adaptation—fine-tuning, RL, distillation—absorbs such interface constraints into model weights, tying solutions to specific checkpoints and training distributions. However, these mismatches typically reflect stable environment-side regularities, motivating adaptation of the runtime harness rather than model weights.

Life-Harness addresses this by adapting the runtime harness—the code and protocol orchestrating agent–environment communication—thus delivering environment-specific, model-agnostic, and reusable solutions across model backbones and held-out evaluation settings (Xu et al., 21 May 2026).

2. Formal Structure of Life-Harness

Life-Harness is formally specified within an episodic LLM agent setting, where each episode is a tuple $\left(x, E, C, B\right)$: $x$ denotes the task, $E$ the deterministic environment (with $\mathtt{Init}$ and $\mathtt{Step}$), $C$ the environment contract (tool schemas, feedback, policies), and $B$ the step budget. Agent policy $\pi_\theta$ is frozen; adaptation proceeds solely through the runtime harness $H$, which is structured into four layers:

1. Environment Contract Layer: Augments the initial contract $C$ by applying $x$0, a set of clarifications mined from training failures (including tool rules, admissible actions, and policies).
2. Procedural Skill Layer: During prompt construction, retrieves relevant, non-parametric skills from memory $x$1 with a retrieval score $x$2 (e.g., using BM25). The top-$x$3 skills $x$4 are inserted into the prompt.
3. Action Realization Layer: Prior to execution, applies deterministic validation to the model’s candidate action $x$5. This yields either $x$6 (proceed with action) or $x$7 (block action with message), depending on schema compliance, type checks, or state guards.
4. Trajectory Regulation Layer: After feedback $x$8, computes $x$9, issuing recovery directives or warnings for loops and no-progress patterns based on available budget.

Harness interventions $E$0 are localized updates (e.g., new guards, contract patches, skill templates), formulated by diagnosing recurrent failure patterns $E$1 in training trajectories. Each intervention is triggered by a condition $E$2 and scored as

$E$3

to measure frequency and mechanistic identifiability within failures (Xu et al., 21 May 2026).

3. Life-Harness Evolution and Algorithmic Workflow

The Life-Harness method evolves through interaction and iterative diagnosis on frozen models:

1. Data Collection: Run agent to gather failed trajectories $E$4 on deterministic environments.
2. Failure Pattern Mining: Diagnose recurring failures $E$5.
3. Intervention Proposal: For each failure $E$6, propose an intervention $E$7 at the earliest layer able to prevent $E$8.
4. Regression Auditing: Ensure proposed interventions do not degrade previously successful cases.
5. Harness Update: Append $E$9 to harness $\mathtt{Init}$0. Iterate until no high-frequency failures persist or saturation is attained.

At runtime, Life-Harness executes according to the following layered protocol (Algorithm 1):

$\mathtt{Init}$3

This architecture strictly prohibits any update to model parameters, preserving model–harness separation and experimental validity in held-out evaluation (Xu et al., 21 May 2026).

4. Experimental Protocol and Benchmarks

Evaluation spans seven deterministic environments, specifically τ-bench (Airline, Retail), τ²-bench (Telecom), and AgentBench (ALFWorld, WebShop, OS, DBBench), each characterized by stable APIs and deterministic transitions. Eighteen LLM backbones were considered—across the Qwen (various sizes), Llama, and xLAM families—covering instruction-tuned, reasoning-specialized, and tool-tuned variants.

Metrics included Pass@1 (single-run success), Pass@3, and Pass^3 (all three runs succeed). Agents were assessed with temperatures fixed at 0.0, and environment step budgets ranging 8–200 per task (Xu et al., 21 May 2026).

5. Empirical Results and Analytical Insights

Life-Harness yielded robust improvements across models and environments. Of 126 model–environment settings, 116 showed improvement with Life-Harness, giving an average relative gain of 88.5% over no-harness baselines.

Aggregate gains by benchmark (averaged over all models):

Benchmark/Setting	Relative Gain (%)
ALFWorld	+84
WebShop	+40
OS	+19
DBBench	+34
Airline Pass@1	+26
Airline Pass^3	+50
Retail Pass@1	+10
Retail Pass^3	+19
Telecom Pass@1	+25
Telecom Pass^3	+27

Harnesses evolved solely from Qwen3-4B-Instruct were applied unmodified to 17 other models, consistently improving 92% of all settings measured. This outcome demonstrates that Life-Harness captures reusable, environment-side structure rather than model-specific artifacts.

Ablation studies show each lifecycle layer is essential: disabling Action Realization causes a −61.7% drop (Airline), and removing Trajectory Regulation leads to a −86.5% drop (ALFWorld). Prompt-only evolution yields modest (10–20%) improvements; full harness adaptation adds 120% relative improvement, evidencing the necessity of runtime interventions.

Life-Harness also complements model-centric adaptation: Qwen2.5-32B+Life-Harness outperforms xLAM-2-32B (tool-specialized) on τ-bench by 7.5pp; further applying Life-Harness on xLAM-2 models yields 6.8–28.9pp gains. Notably, tool-specialized model training can degrade out-of-distribution generalization, whereas interface harnessing remains robust (Xu et al., 21 May 2026).

6. Conceptual Implications and Future Directions

Empirical findings substantiate that major agentic failure modes in deterministic domains originate from the interface boundary rather than model weights. By evolving a structured, four-layered runtime harness, environment-invariant knowledge (contracts, skills, recovery logic) is operationalized in a model-agnostic manner.

A plausible implication is that runtime interface adaptation and model-centric training should be regarded as complementary strategies: model training adapts $\mathtt{Init}$1 to a distribution (with distribution-specific and model-specific gains), whereas Life-Harness adapts $\mathtt{Init}$2 to the environment, delivering performance transferable across LLM families.

Open research directions include: extending the approach to stochastic, open-ended, or multi-agent domains; automating harness evolution via integration with trajectory diagnostics, symbolic failure classifiers, and intervention synthesis; and exploring joint optimization of model weights and interface harnesses for further gains.

7. Resources and Availability

Code and a comprehensive harness inventory for Life-Harness, along with experimental artifacts, are publicly released at [https://github.com/Tianshi-Xu/Life-Harness]. The framework offers a reproducible paradigm for runtime interface adaptation, facilitating robust, cross-model improvement in LLM-agent systems (Xu et al., 21 May 2026).