💬 NLP

Co-Harness: Co-Evolving Harnesses and Model Weights for LLM Agents

Co-Harness is a framework that jointly optimizes an LLM agent's runtime harness and model parameters through an alternating process of harness refinement and model fine-tuning, enabling autonomous recovery, efficiency gains, and strategy discovery beyond the limitations of fixed-harness post-training.

Original authors: Zhengyu Chen, Teng Xiao, Huaisheng Zhu, Yige Yuan, Luan Zhang, Jingang Wang

Published 2026-07-28

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhengyu Chen, Teng Xiao, Huaisheng Zhu, Yige Yuan, Luan Zhang, Jingang Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a brilliant but inexperienced apprentice how to solve complex puzzles. In the world of artificial intelligence, this apprentice is a "Large Language Model" (LLM)—a computer program that can read, write, and reason. But an LLM doesn't work in a vacuum; it needs a workshop. This workshop is called a "Harness." Think of the Harness as the set of tools, the instruction manual, the safety rails, and the memory bank the apprentice uses to do their job. It includes things like how the computer talks to a calculator, how it remembers previous steps, and how it knows when to ask for help.

For a long time, scientists treated the apprentice and the workshop as two separate things. They would spend years training the apprentice to be smarter, but they would keep the workshop exactly the same, using the same old tools and instructions they had designed by hand. The problem is, if the workshop has a wobbly table or a confusing map, even the smartest apprentice will stumble. They might fail not because they aren't smart enough, but because the environment they are working in is broken. This paper asks a simple but powerful question: What if we didn't just train the apprentice, but also fixed the workshop at the same time? What if the apprentice could help us figure out what was wrong with the tools, and then we could teach them using the new, better tools?

This is exactly what the researchers behind Co-Harness discovered. They created a system where the AI model and its "workshop" (the Harness) evolve together, like a dance partner learning to move in sync. Instead of just fixing the AI and hoping for the best, their system looks at every time the AI fails a task. It asks, "Was this because the AI was confused, or because the tool it was holding was broken?" If the tool was the problem—say, a confusing instruction or a missing safety hook—the system automatically fixes that tool. Then, it uses the AI's success with the new tool to teach the AI how to be even better.

The results are like watching a snowball grow into a boulder. The researchers tested this on tough math and logic puzzles. They found that by letting the AI and its tools improve together, the system got significantly smarter. After just two rounds of this "co-evolution," the AI's accuracy on difficult math tests jumped by over 20 percentage points on average. In one specific case, the system ran for over 200 hours without any human help, fixing its own crashes, speeding up its work, and even inventing a new strategy of combining multiple attempts to get the right answer. The paper suggests that this "dual-loop" approach—fixing the environment and training the brain simultaneously—is a much more powerful way to build intelligent agents than the old method of keeping the environment static. It turns out that to make a genius, you sometimes have to build a better playground first.

Technical Summary: Co-Harness: Co-Evolving Harnesses and Model Weights for LLM Agents

1. Problem Statement

Post-training agents for automated AI research traditionally involves optimizing model parameters while treating the runtime system—the "Harness"—as a fixed, static scaffold. The Harness encompasses prompts, tools, skills, middleware, memory policies, and retry logic. Current pipelines create a fundamental mismatch: the model is updated via supervised fine-tuning (SFT) or reinforcement learning based on trajectories generated by a static Harness, yet the Harness itself is never optimized based on the failures observed during training.

This asymmetry leads to a bottleneck where suboptimal scaffolding (e.g., ambiguous prompts, incorrect tool schemas, missing retry hooks, or memory overflow) prevents the generation of high-quality training trajectories. Consequently, the model cannot learn to overcome these environmental limitations, and the fixed Harness becomes a ceiling for agent performance, particularly in complex Tool-Integrated Reasoning (TIR) settings where a single schema error can abort an entire multi-turn session.

2. Methodology: Co-Harness Framework

The authors propose Co-Harness, a dual-loop framework that jointly optimizes the agent Harness ( $\phi$ ) and model parameters ( $\theta$ ) through an alternating process. The core hypothesis is that a better Harness produces cleaner, more informative trajectories, which train a stronger model; this stronger model, in turn, exposes higher-order Harness bottlenecks that were previously invisible, enabling further scaffolding improvements.

The framework operates in two alternating loops per round $t$ :

A. The Co-Harness Loop (Harness Optimization)

In this phase, the model $\theta_t$ is held fixed while the Harness $\phi_t$ is evolved.

Failure Collection: The agent executes tasks under $(\theta_t, \phi_t)$ , collecting a set of failed trajectories $F^-_t$ .
HarnessCritic Analysis: An LLM-based component, HarnessCritic ( $C$ ), analyzes these failures. It attributes each failure to a structured root cause within the Harness configuration (e.g., prompt ambiguity, tool schema error, missing skill, middleware mismatch, or memory overflow).
Diff Generation & Validation: HarnessCritic proposes local patches (differences) to the Harness. These patches are validated by running "rollout" tests to ensure they improve the targeted failure mode without causing regression in held-out behaviors.
Registry Update: Validated patches are committed to a versioned Harness registry, updating the active configuration to $\phi^*_t$ .

B. The Model Alignment Loop (Model Optimization)

In this phase, the evolved Harness $\phi^*_t$ is held fixed while the model is updated.

High-Quality Trajectory Collection: The current model $\theta_t$ is run under the improved Harness $\phi^*_t$ to generate a new dataset $D_t$ of high-quality trajectories.
Supervised Fine-Tuning (SFT): The model is fine-tuned on $D_t$ to internalize the behaviors enabled by the improved scaffolding, resulting in an updated model $\theta_{t+1}$ .
Iteration: The new model $\theta_{t+1}$ and the evolved Harness $\phi^*_t$ become the starting point for the next round, where the stronger model may reveal new, more complex Harness limitations.

Failure Attribution Taxonomy

HarnessCritic classifies failures into five actionable Harness dimensions and one abstention label:

Prompt Ambiguity ( $P$ ): Instruction underspecification or conflicting goals.
Tool Schema Error ( $T$ ): Invalid tool calls, bad argument schemas, or backend mismatches.
Skill Missing ( $S$ ): Absence of reusable routines or decomposition primitives.
Middleware Mismatch ($Mid$): Flawed loop protocols, hook behaviors, or context management logic.
Memory Overflow ( $M$ ): Persistent state overflow, stale memory, or retrieval failures.
Agent Error: Cases where the failure is intrinsic to the model's reasoning rather than the Harness.

3. Key Contributions

Joint Optimization Formulation: The paper formulates agent post-training as a joint optimization problem over model weights and the runtime Harness, moving beyond the paradigm of treating prompts and tools as fixed infrastructure.
HarnessCritic: The introduction of a failure-driven procedure that performs structured attribution of errors to specific Harness components and proposes validated, local diffs for repair.
Compounding Gains Mechanism: The demonstration that alternating Harness evolution and Model Alignment creates a positive feedback loop, where improvements in scaffolding lead to better training data, which yields stronger models capable of leveraging even more sophisticated scaffolding.

4. Experimental Results

The framework was evaluated on Tool-Integrated Reasoning (TIR) tasks using mathematical benchmarks (AIME 2024, AIME 2025, HMMT 2025) and two model scales (Qwen3-8B and Qwen3-32B).

Performance Gains: Two rounds of Co-Harness co-evolution resulted in an average accuracy improvement of +20.4 percentage points (pp) across benchmarks and model scales compared to the baseline (a Harness evolved once without model SFT).
- The largest gain was observed on the most difficult benchmark (HMMT 2025) for the Qwen3-32B model, with an improvement of +27.2 pp.
- The system surpassed carefully hand-designed, static Harness configurations by an average of +24.7 pp.
Autonomous Case Study: In a 200+ hour autonomous run on AIME 2024, the system:
- Recovered from an initial system crash (vllm KV cache error).
- Improved inference efficiency by 8.7× (reducing wall-clock time from 3.78h to 1.11h) through global-batched inference.
- Discovered an ensemble strategy (majority voting of diverse seeds) that achieved the best accuracy (63.3%) within the time budget, all without human intervention.
Robustness: The system successfully detected and rolled back regressions (e.g., domain-specific prompts that conflicted with the model's internal reasoning) using its version-controlled registry.

5. Significance and Claims

The paper argues that Co-Harness addresses a critical limitation in current agent development: the decoupling of model training from the environment that generates training data. By treating the Harness as a learnable variable rather than a static scaffold, Co-Harness enables compounding self-improvement.

The authors claim that this approach realizes a "positive feedback spiral" where:

A better Harness produces cleaner tool use and fewer avoidable crashes.
These cleaner trajectories train a stronger model.
The stronger model can exploit richer, more complex Harness configurations that were previously unusable.

This contrasts with prior work that either optimizes the model for a fixed Harness (standard SFT/RL) or optimizes the Harness for a fixed model (inference-time search). Co-Harness is presented as the first framework to co-evolve both, suggesting that the interdependence of scaffolding and model weights is a primary axis for improving agent capabilities beyond fixed-harness post-training.

Limitations noted by the authors: The approach requires a capable critic LLM, a sufficient volume of failure trajectories for cold-start, and significant compute for multi-round SFT. Additionally, attribution accuracy can degrade for failure modes requiring complex counterfactual reasoning, and structural redesigns of the Harness remain a human responsibility.