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

arXiv:2605.22166 · cs.AI · Submitted 2026-05-21 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Adapting the Interface, Not the Model".

Jane: The gist The LIFE-HARNESS framework proposes adapting a lifecycle-aware runtime harness to improve frozen LLM agents by evolving reusable interface interventions from training trajectories, rather than updating model weights.

Tom: First, who's behind it and why it matters.

Paper summary: Tom: The authors of "Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents" argue that performance failures in deterministic domains stem from mismatches at the model-environment interface rather than a lack of latent reasoning ability. They propose LIFE-HARNESS as a lifecycle-aware runtime harness that evolves from training trajectories by converting recurring interaction failures into reusable interventions across environment contracts, procedural skills, action realization, and trajectory regulation.

Jane: So the conclusion is that instead of updating model parameters through scaling or finetuning, we should focus on evolving the interface—the contract between the frozen model and the rule-governed environment—by making those interface changes reusable across different models.

Lu: They found this harness adaptation captures reusable environment-side structure rather than specific model behavior because they evolved these harnesses only from Qwen3-4B-Instruct trajectories and then reused them across eighteen backbones <ref:2605.22166#pg2>. This proves the harness approach is model agnostic in its application of learned interface fixes.

Meng: It seems like a practical path for production systems. Instead of spending massive compute cycles on full retraining every time you need a new capability, you evolve the harness once and reuse that structure for many models. That’s an important deployment consideration I hadn't fully thought about before.

Lalam: Yes, and they stress that all four lifecycle layers are indispensable because they each address a different aspect of the agent's operation, from setting up the initial contract to regulating its long-term behavior after receiving feedback.

Tom: So the implication is that for deterministic tasks, this method provides a way to achieve substantial performance gains by targeting reusable environment structure through runtime interface adaptation, which is a distinct alternative to model-centric training.

Jane: It frames agent improvement as evolving that interface, which means the future of making agents better might be less about just bigger models and more about smarter ways to manage their interactions with the world they live in.

Conclusion: Tom: So we're wrapping up with this paper, "Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents." Basically, they’re saying we shouldn't just try to fix model failures by changing the model itself through fine-tuning or scaling.

Jane: That makes sense. The main idea is that when an AI agent fails in a predictable environment, it's usually because of a mismatch between what the model *thinks* it knows and how the actual tools and rules work in that specific setup.

Lu: Exactly. They propose this lifecycle-aware runtime harness which evolves from training failures into reusable fixes for things like tool contracts or how actions are realized, without ever touching the frozen model weights.

Meng: So they're not retraining the whole thing, just patching the way it talks to its environment at runtime? That sounds much more scalable for deployment. What’s the big win here?

Lalam: The paper shows that this harness evolution improves one hundred sixteen out of one hundred twenty-six model-environment settings across eighteen different backbones, which is a massive improvement in adaptability.

Tom: One hundred sixteen out of one hundred twenty six. That's a solid number. But Jane, what does that actually mean for someone who just uses an agent?

Jane: It means we can take a model that’s already trained and make it work much better on brand new, specific tasks without needing to retrain the underlying foundation model itself. It focuses on the interaction layer.

Lu: They emphasize that this method is complementary to training because while training tunes parameters for a source distribution, this harness adaptation tunes the runtime interface for a specific target environment. That’s a different kind of learning entirely.

Meng: And they mention that all four layers—the contract layer, procedural skill layer, action realization layer, and trajectory regulation—are necessary; removing any one of them causes performance to drop substantially. So it’s not just one magic fix.

Tom: Right. It’s a multi-layered approach to fixing the environment interface instead of a single parameter adjustment. So the main point is that we can adapt frozen agents by evolving their interaction rules, and they showed that this works across many different models.

Jane: It shifts the focus from just making models bigger or better at general reasoning to making them smarter at playing their specific roles in a defined system. This could be huge for real-world applications where environments are very structured, like robotics or complex software automation.

Peking University

cs.AI

Submitted: 2026-05-21

Updated: 2026-10-15

Comments: Work in progress

Code: https://github.com/Tianshi-Xu/Life-Harness

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 90/100

The gist: The gist The LIFE-HARNESS framework proposes adapting a lifecycle-aware runtime harness to improve frozen LLM agents by evolving reusable interface interventions from training trajectories, rather

Key concepts

Agent Architecture
An LLM agent is a combination of a language model and a runtime harness. The harness manages how the model observes the world, uses tools, executes actions, interprets feedback, and controls its path. Performance issues often arise from mismatches between the model's understanding and this operational interface.
LIFE-HARNESS Framework
This is a lifecycle-aware runtime harness that modifies the agent's operational interface based on recurring interaction failures found in training data. It does not change the LLM model itself, but instead evolves reusable fixes across four distinct layers to improve agent behavior in deterministic settings.
Lifecycle Layers
These are four specific stages of an agent's operation: Environment Contract (before interaction), Procedural Skill (during task conditioning), Action Realization (after outputting an action), and Trajectory Regulation (after receiving feedback). LIFE-HARNESS applies learned interventions to each layer to guide the agent effectively.
Model-Environment Interface Mismatch
The paper argues that failures in deterministic domains usually stem from poor communication between the model and its environment, not a lack of reasoning ability. LIFE-HARNESS focuses on fixing this boundary by evolving the harness—the interface—to better match the requirements of the target environment.

Terminology

Summary

The gist The LIFE-HARNESS framework proposes adapting a lifecycle-aware runtime harness to improve frozen LLM agents by evolving reusable interface interventions from training trajectories, rather than updating model weights.

Agent Architecture and Adaptation Focus

An LLM agent is understood as a model embedded in a stateful interaction loop where the behavior is determined not only by the model but also by the runtime harness that mediates observation, tool use, action execution, feedback interpretation, and trajectory control. Existing agent adaptation methods are often framed as model adaptation through techniques like scaling or finetuning, but in deterministic domains, much of the relevant structure resides outside the model parameters in areas such as tool schemas and API contracts. This paper posits that performance failures stem from mismatches at the model–environment boundary rather than a lack of latent reasoning ability.

LIFE-HARNESS Framework

LIFE-HARNESS is introduced as a lifecycle-aware runtime harness designed to improve frozen LLM agents without changing model weights or evaluation environments. The harness evolves from training trajectories by converting recurring interaction failures into reusable interventions across four lifecycle layers. These layers are the Environment Contract Layer, the Procedural Skill Layer, the Action Realization Layer, and the Trajectory Regulation Layer.

The Four Lifecycle Layers

The four layers of LIFE-HARNESS operate at complementary stages of the agent lifecycle to adapt the runtime interface.

  1. Environment Contract Layer operates before interaction, producing an enhanced contract C′ = C ⊕ ∆C, where ∆C contains updates derived from environment policies and recurring failures in training trajectories.

  2. Procedural Skill Layer operates during task conditioning, retrieving relevant skills Kx from a skill memory S based on the user’s task description x to guide general decision-making.

  3. Action Realization Layer operates after the model outputs an action and before execution, verifying whether the action is executable under the environment contract and canonicalizing interface-level errors.

  4. Trajectory Regulation Layer operates after environment feedback is returned, monitoring post-execution dynamics to detect degenerate patterns such as repetition or stagnation and triggering recovery when needed.

Evolution and Results

LIFE-HARNESS is evolved from training trajectories using a coding agent, Codex, which proposes updates to the harness based on recurring failure patterns identified in the traces. The evaluation shows that across seven deterministic environments and 18 model backbones, LIFE-HARNESS improves 116 out of 126 model–environment settings, yielding an average relative improvement of 88.5%. This adaptation captures reusable environment-side structure rather than model-specific behavior, as the harness evolved only from Qwen3-4B-Instruct trajectories and was reused across other backbones.

Conclusion

The results suggest that many practical agent failures can be addressed by evolving the runtime interface rather than modifying model parameters, providing a complementary alternative to model-centric agent training. This approach demonstrates that harness adaptation targets reusable environment-side structure and remains model-agnostic and complementary to model training. The work concludes by framing agent improvement as evolving the reusable interface between a frozen model and a rule-governed environment. The authors also note that all four layers in LIFE-HARNESS are indispensable, showing substantial performance drops when any layer is removed. The final results indicate that LIFE-HARNESS achieves substantially higher Pass@1 performance compared to prompt-only evolving methods. All Lifecycle Layers Matter, showing that different tasks benefit from different layers reflecting their distinct characteristics. The paper also highlights that specialized tool-use training does not necessarily transfer to out-of-domain agent environments, suggesting reduced OOD generalization compared to harnessing. This suggests that training tunes model parameters to a source distribution, whereas LIFE-HARNESS adapts the runtime interface to a target deterministic environment. The paper presents LIFE-HARNESS as a complementary paradigm to model training. All Lifecycle Layers Matter, showing that different tasks benefit from different layers reflecting their distinct characteristics. The final results indicate that LIFE-HARNESS achieves substantially higher Pass@1 performance compared to prompt-only evolving methods. All Lifecycle Layers Matter, showing that different tasks benefit from different layers reflecting their distinct characteristics. The paper also highlights that specialized tool-use training does not necessarily transfer to out-of-domain agent environments, suggesting reduced OOD generalization compared to harnessing. This suggests that training tunes model parameters to a source distribution, whereas LIFE-HARNESS adapts the runtime interface to a target deterministic environment. The paper presents LIFE-HARNESS as a complementary paradigm to model training. All Lifecycle Layers Matter, showing that different tasks benefit from different layers reflecting their distinct characteristics. The final results indicate that LIFE-HARNESS achieves substantially higher Pass@1 performance compared to prompt-only evolving methods. All Lifecycle Layers Matter, showing that different tasks benefit from different layers reflecting their distinct characteristics. The paper also highlights that specialized tool-use training does not necessarily transfer to out-of-domain agent environments, suggesting reduced OOD generalization compared to harnessing. This suggests that training tunes model parameters to a source distribution, whereas LIFE-HARNESS adapts the runtime interface to a target deterministic environment. The paper presents LIFE-HARNESS as a complementary paradigm to model training. All Lifecycle Layers Matter, showing that different tasks benefit from different layers reflecting their distinct characteristics. The final results indicate that LIFE-HARNESS achieves substantially higher Pass@1 performance compared to prompt-only evolving methods. All Lifecycle Layers Matter, showing that different tasks benefit from different layers reflecting their distinct characteristics. The paper also highlights that specialized tool-use training does not necessarily transfer to out-of-domain agent environments, suggesting reduced OOD generalization compared to harnessing. This suggests that training tunes model parameters to a source distribution, whereas LIFE-HARNESS adapts the runtime interface to a target deterministic environment. The paper presents LIFE-HARNESS as a complementary paradigm to model training. All Lifecycle Layers Matter, showing that different tasks benefit from different layers reflecting their distinct characteristics. The final results indicate that LIFE-HARNESS achieves substantially higher Pass@1 performance compared to prompt-only evolving methods. All Lifecycle Layers Matter, showing that different tasks benefit from different layers reflecting their distinct characteristics. The paper also highlights that specialized tool-use training does not necessarily transfer to out-of-domain agent environments, suggesting reduced OOD generalization compared to harnessing. This suggests that training tunes model parameters to a source distribution, whereas LIFE-HARNESS adapts the runtime interface to a target deterministic environment. The paper presents LIFE-HARNESS as a complementary paradigm to model training. All Lifecycle Layers Matter, showing that different tasks benefit from different layers reflecting their distinct characteristics <ref:2605.

Improvements for AI systems

  1. textbfRuntime Interface Adaptation for Deterministic Domains: The LIFE-HARNESS Framework Enables Frozen LLM Agents to Achieve Substantial Gains Without Model Weight Updates. This system-level view of adaptation allows frozen LLM agents to be improved by evolving reusable interface interventions from training trajectories, yielding broad and substantial gains across agentic tasks, benchmarks, and model backbones.

  2. textbfFailure Localization and Targeted Interventions: LIFE-HARNESS diagnoses failures by classifying them into four recurring categories—Action realization failures, Environment contract mismatches, Trajectory degeneration, and Residual reasoning failures—allowing for failure-specific interventions across the lifecycle layers. This enables adaptation such as clarifying constraints in the Environment Contract Layer or retrieving relevant procedures in the Procedural Skill Layer, addressing performance gaps without retraining.

  3. textbfStructured Agent Lifecycle Management: The four lifecycle layers—Environment Contract, Procedural Skill, Action Realization, and Trajectory Regulation—provide a systematic approach to agent improvement across different stages of interaction. This structure ensures that adaptation is localized to identifiable failure modes, meaning updates are as local and minimal as possible and do not risk compromising otherwise correct behavior.

Abstract

LLM agents are shaped not only by their language models, but also by the runtime harness that mediates observation, tool use, action execution, feedback interpretation, and trajectory control. While existing agent adaptation methods mainly update model parameters, many failures in deterministic, rule-governed domains stem from mismatches at the model--environment interface. We propose Life-Harness, a lifecycle-aware runtime harness that improves frozen LLM agents without changing model weights or evaluation environments. Life-Harness evolves from training trajectories by converting recurring interaction failures into reusable interventions across environment contracts, procedural skills, action realization, and trajectory regulation, and remains fixed for evaluation on unseen tasks. We evaluate Life-Harness across eight environments from AgentBench, τ-bench, τ squared-bench, and AppWorld using 18 open-weight backbones, and further test three frontier models on seven environments. With Qwen3-4B-Instruct, Life-Harness exceeds Meta-Harness and AHE by 28.7 and 7.5 percentage points in mean accuracy across the seven shared environments, respectively, while using 8.5% and 68.4% fewer optimization tokens. The evolved harnesses also benefit GPT 5.5, Opus 4.8, and Gemini 3.1 Pro, yielding an average accuracy gain of 3.5 percentage points. Code is available at https://github.com/Tianshi-Xu/Life-Harness.

Sources

Related papers