Every LLM-based agent today is shaped by two things: the model itself and the harness around it. The harness is the runtime scaffolding that manages tools, formats observations, maintains conversation state, enforces permissions, and orchestrates multi-step workflows. Recent work has shown that harness design can change an agent's end-to-end performance by as much as six times on a fixed model. That raises an obvious question: why not just bake the best harness's behavior directly into the model, so the specialized scaffolding can be discarded?
Haoran Ye, Yuxing Lu, Haonan Dong, Zhaochen Su, and Guojie Song tackle exactly this question in a paper submitted to arXiv in September 2026. They call their approach Harness-Zero, and the core idea is harness distillation: taking the behavior induced by a domain- or instance-optimized external harness and transferring it into model weights through supervised fine-tuning, so that the gains survive when the harness is removed at deployment. The results are striking, with macro-average task success improving from 23.3 percent to 44.3 percent after distillation, a figure that exceeds the 41.7 percent the base model achieves even when the original optimized harness is still attached.
The Harness Attachment Problem
An agent harness is not a prompt or a single tool. It is the entire runtime environment that mediates between the model and the external world. It decides what observations the model sees, what actions are available, how tool calls are parsed and executed, and how feedback flows back into the conversation. When a harness is well-designed for a specific domain, it can dramatically improve agent performance by providing the right structure at the right time.
The problem is that these gains are hostage to the harness itself. The best harness for a coding task differs from the best harness for a scientific analysis task, which differs again from the best harness for a knowledge-work workflow. Even within a single domain, the optimal harness can vary across instances and models. A general-purpose agent therefore faces an unenviable choice: either use a suboptimal shared harness that compromises across all tasks, or route among an ever-growing library of specialized harnesses, each tuned to a specific context.
The second option creates operational complexity. Each specialized harness must be maintained, tested, and kept compatible with the underlying model as it evolves. The first option leaves performance on the table. Neither solution scales.
The Core Challenge: Two Harnesses, Two Action Spaces
Harness distillation sounds simple in principle: train a model using the optimized harness as a teacher, then deploy the model with a simpler harness. In practice, the problem is that the two harnesses operate with different action spaces and different available information. The optimized harness might present the model with a rich set of tools, structured observations, and specific formatting conventions. The target harness at deployment, by contrast, might have fewer tools, different observation formats, or a different execution loop.
Guidance from the optimized harness therefore cannot serve directly as supervision for the target model. A teacher demonstration generated under the rich harness uses actions and observation formats that the student model, running under the sparser harness, cannot replicate. The distillation pipeline must translate between these two worlds.
How Harness-Zero Works
Harness-Zero solves the translation problem with a clever architectural choice: it inserts a harnessing agent between the optimized harness and the student model. This harnessing agent is itself an LLM that is guided by the optimized harness during training. Its job is to take student responses that would be generated under the target harness, correct them to align with what the optimized harness expects, and then execute those corrected responses in the target harness's action space.
The process works as follows. The student model generates a response under the target harness's constraints. The harnessing agent, which has been trained to understand the optimized harness's expectations, reviews that response and corrects it where necessary. The corrected response is then executed in the target harness, producing a trajectory that pairs the student model's original output with the harness-guided correction. These corrected trajectories become the training data for fine-tuning the student model.
Through this mechanism, harness guidance is converted into training demonstrations. The student model learns not from raw trajectories generated by the optimized harness, but from corrected versions of its own outputs that incorporate the optimized harness's behavioral patterns. Over many such iterations, the model internalizes the harness-induced behaviors directly into its weights.
The critical property of this approach is that once fine-tuning is complete, the specialized optimized harness can be removed entirely. The student model now carries the distilled behaviors in its parameters, so it performs at the same level under a simple fixed target harness as it did under the complex optimized one.
The Three-Stage Pipeline
The paper evaluates across three domains: knowledge work, tool use, and science. Each domain presents a different flavor of the harness distillation problem. Knowledge work tasks require the model to follow complex multi-step workflows with structured outputs. Tool use tasks require precise API calls and environment interactions. Science tasks demand rigorous reasoning and domain-specific tool chains.
The pipeline spans three stages. First, an optimized harness is trained or evolved for a specific domain or instance. This harness serves as the teacher, providing behavioral guidance that shapes how the student should act. Second, the harnessing agent operates in the target harness's action space, correcting student responses and generating training trajectories. Third, the student model is fine-tuned on these trajectories, internalizing the distilled behaviors.
A distinguishing feature of the approach is that the harnessing agent itself can be a frontier LLM. This is significant because it means the distillation process leverages the capabilities of a powerful model to teach a weaker model, rather than requiring the distilled behaviors to be explicitly programmed as rules or templates.
Results That Matter
The numerical results are the most compelling part of the paper. Across knowledge work, tool use, and science domains, the findings are consistent and substantial.
First, agent-as-harness outperforms code-as-harness when both use the same evolved harness. This comparison isolates the effect of the distillation mechanism: rather than hard-coding harness behavior as a fixed set of rules, the approach learns it from examples. The result is a more flexible and generalizable distilled behavior that adapts better to the target harness's constraints.
Second, the headline number deserves careful attention. The base model, without any harness distillation, achieves a macro-average task success of 23.3 percent across the evaluated domains. With the optimized harness still attached, it reaches 41.7 percent. After Harness-Zero distillation and removal of the specialized harness, the model achieves 44.3 percent. This means the distilled model outperforms the base model even when that base model has the full optimized harness available. The distillation did not just recover the harness's gains; it exceeded them, suggesting that the fine-tuning process also improved the model's intrinsic capabilities in ways that complement the distilled behaviors.
Third, the recovery rate for behaviors absent from the base model is 82.3 percent on average, across 28 distinct patterns in the three domains. These 28 patterns represent specific behavioral primitives that the optimized harness induces but that the base model does not exhibit naturally, such as particular tool-calling sequences, verification steps, or structured output formats. Recovering 82.3 percent of these patterns indicates that the distillation is genuinely transferring specialized behaviors rather than simply amplifying what the model already knows.
What Makes This Different from Prior Work
The paper occupies a unique position in a rapidly growing field. Prior work on agent distillation has focused on transferring reasoning trajectories from strong teacher models to weaker student models, as in Agent Distillation (Nardien et al., 2025), which distills tool-using agent behavior into small models through reason-act-observe trajectories. Other approaches have focused on distilling Model-Context Protocols (MCPs) as reusable task-solving modules, or on training-free methods that directly reuse teacher-generated tool calls.
What distinguishes Harness-Zero is its specific focus on the harness layer. Prior distillation methods transfer model knowledge and reasoning patterns. Harness-Zero transfers the runtime behaviors and interaction patterns that are normally the exclusive province of the harness. This is a fundamentally different type of knowledge: not what the model knows, but how it should act within a structured interaction loop.
Related work on harness improvement, such as Self-Harness (2026), which enables an LLM to improve its own operating harness, and LIFE-HARNESS (2026), which evolves runtime interfaces without updating model weights, shares the same conceptual space but approaches it from a different direction. Those works modify the harness around a fixed model. Harness-Zero modifies the model around a fixed harness, and then discards the harness.
The security-oriented Agent Harness Distillation work (Cui et al., 2026) studies the risks of extracting inference-time harnesses from agent systems. Harness-Zero is not concerned with security; it is concerned with efficiency. But both lines of work recognize that harnesses carry substantial, transferable knowledge that goes beyond what the base model encodes.
Limitations and Honest Inferences
The approach inherits several constraints worth acknowledging. First, the harnessing agent must be capable of understanding the optimized harness's expectations well enough to correct student responses. If the optimized harness induces behaviors that are difficult to articulate or correct, the distillation pipeline may lose fidelity. The 82.3 percent recovery rate leaves room for improvement, and the remaining 17.7 percent of unrecovered patterns may represent behaviors that are highly context-dependent or difficult to transfer through this mechanism.
Second, the fine-tuning process requires labeled trajectories generated through the harnessing agent's corrections. This adds a training overhead that is not trivial, even though it is a one-time cost amortized across deployment. The paper does not report the computational cost of generating these training trajectories, which is an important practical consideration.
Third, the evaluation is limited to three domains and a specific set of 28 behavioral patterns. The generalizability to other domains, particularly those with very different interaction structures, is not established. A harness optimized for a creative writing workflow might induce patterns that are harder to distill than the tool-calling and verification patterns studied here.
Fourth, the approach assumes that the base model is capable of learning the distilled behaviors through fine-tuning. If the model's architecture or training data place hard limits on what behavioral patterns it can acquire, distillation may not be able to overcome those limits.
Fifth, the paper does not fully address the question of how the distilled behaviors interact with the model's original capabilities. The 44.3 percent macro-average success rate exceeds both the base model and the base model with the harness attached, but it is not clear how much of that gain comes from genuine distillation versus from the fine-tuning process itself improving the model on the specific tasks evaluated.
What This Means for Practitioners
For teams deploying LLM agents in production, the practical implication is significant. Today, building a capable agent requires not only selecting a model but also engineering a harness tailored to the task domain. Harness-Zero suggests a different deployment model: invest once in training a specialized harness, use it to generate corrected training data, fine-tune the model, and then retire the harness. The resulting model is simpler to deploy, easier to maintain, and does not require ongoing harness engineering.
This is particularly relevant as the number of agent frameworks and harness designs continues to grow. Teams that currently maintain separate harness configurations for different tasks could, in principle, distill those harness behaviors into shared models, reducing operational complexity. The approach also has implications for edge deployment: a distilled model that carries harness-like behaviors in its weights requires less runtime infrastructure than a model that depends on a complex external harness.
The privacy angle is also worth noting. Because the harness is removed at deployment, there is no external runtime system mediating between the model and the environment. The model itself carries the interaction patterns. This reduces the attack surface and simplifies compliance with data governance requirements.
The Bigger Picture
Harness-Zero represents a conceptual shift in how the agent community thinks about the relationship between models and their surrounding infrastructure. For years, the implicit assumption has been that the harness and the model are complementary and must coexist. The paper challenges this assumption by arguing that the harness's most valuable contributions can be internalized into the model, making the specialized harness redundant.
This is not the first paper to question the model-harness boundary. LIFE-HARNESS adapts the runtime interface without touching model weights. Self-Harness lets the model improve its own harness. HarnessBridge learns a bidirectional controller. But Harness-Zero goes further than any of these by proposing that the boundary itself can be dissolved: the model becomes the harness, absorbing its external scaffolding into its own parameters.
The 23.3 percent to 44.3 percent improvement, combined with the 82.3 percent recovery rate across 28 behavioral patterns, provides a quantitative foundation for this vision. The approach works, it works across domains, and the gains are large enough to justify the engineering investment. The question for the field is no longer whether harness distillation is possible, but how far it can be pushed toward general-purpose agents that carry their own harnesses in their weights.