HomeArticle

The covert war in the second half of the Agent takes place at the Harness scheduling layer.

晓曦2026-07-10 14:34
The value of domestic models is shifting from single-point capabilities to integrated capabilities.

Article by Wang Yi

In a recent evaluation conducted by DRACO, a set of data has drawn industry attention.

This is an open benchmark test consisting of 100 highly challenging research and analysis tasks. The open-source Agent framework OpenSquilla 0.5.0 organizes four domestic models—DeepSeek, GLM, Kimi, and Qwen—into a parallel proposal architecture, with one model aggregating the final output.

The results show that this all-domestic lineup achieved a quality score of 60.85, slightly higher than the 59.80 of the latest flagship model Fable5. The average cost per task is approximately 1/3 of the latter, at $0.39 and $1.21 respectively.

In another comparison, this multi-model integration solution delivers a higher quality score than running Claude Opus 4.8 and GPT-5.5 individually, while reducing costs by about 86% to 92%.

The real question raised by this set of data is whether complex tasks still need to be defaulted to the single most powerful model.

Over the past two years, competition among large models has primarily centered on foundational models, with model capabilities, context window length, and benchmark performance receiving the most attention. Moving to the application layer, a task often goes through multiple rounds of execution. How models are selected, how they collaborate, how results are verified, and how failures are recovered all affect the final quality and cost.

The model itself is certainly important, but how models are organized is also becoming equally critical.

01. The Overlooked Harness

To understand this shift, we need to break down the basic structure of an Agent. The industry commonly uses the formula to express the relationship between Agent and Harness — Agent = Base Models + Harness.

This means the foundational models provide capabilities such as language understanding, reasoning, code generation, and tool usage, while the Harness is responsible for integrating these capabilities into specific workflows to make outputs executable, observable, and verifiable.

If models provide capabilities, the Harness governs how those capabilities are used. It includes context management, execution control, tool invocation, output verification, failure recovery, and also the collaboration arrangements between multiple models. When a single model is called in isolation, it only answers questions; once integrated into a Harness, it becomes part of a complete workflow.

The value of the Harness goes far beyond mere engineering wrapping.

The Primitive Rhythm team previously noted in their Claw-SWE-Bench research that for the same model, task completion rates can differ by up to 27.4 percentage points under different Harness configurations. In other words, without changing the model at all, simply adjusting the peripheral execution framework can lead to a difference of nearly 30% in final results.

In everyday scenarios, this gap becomes even more pronounced.

In many tasks, model invocation rarely stops at a single question-and-answer exchange. The system needs to understand the task, break down steps, retrieve information, call tools, and then complete verification and correction. Instability in any intermediate link will affect the final delivery.

Previously, simply assigning all tasks to the most expensive model was a hassle-free choice. To avoid failures in complex tasks, people would default to calling flagship models, while simple tasks were mixed in, consuming high-cost computing power just the same. Bills kept rising, yet results remained not fully controllable.

As the number of models increased, selection criteria also changed. Different models began to develop their own areas of focus. The challenge is no longer just about choosing the strongest model, but also matching the right model to each specific step.

This is exactly what the Harness is designed to do. It does not participate in foundational model competition, but it determines how models are scheduled, complement each other, and are verified — ultimately deciding whether stable delivery can be achieved.

02. OpenSquilla 0.5.0: Organizing Models Into a Team

The OpenSquilla 0.5.0 Preview and the concurrently released "Agentic Routing" technical report apply this philosophy to a concrete product.

OpenSquilla is not a new large language model — it is an open-source, locally deployable Agent runtime framework. It does not train foundational models; instead, it focuses on scheduling and orchestration above the model layer.

Agentic Routing can be understood as step-level model routing. Traditional routing usually makes only one judgment when a request arrives, selecting a single model based on the user's question. Agentic Routing, by contrast, makes judgments during the Agent's execution process, reselecting the appropriate model at each step based on the current state.

Simple tool calls can be handled by lightweight models; complex reasoning, long-context summarization, or code generation may require more powerful models. Some steps even invoke multiple models simultaneously to generate solutions in parallel, after which the system aggregates the outputs.

The core capability of version 0.5.0 is multi-model integrated collaboration.

In this mechanism, the four domestic models are placed in parallel proposal positions, independently completing search, reasoning, and answer generation. A fifth model takes on the aggregation role, synthesizing multiple candidate answers to produce the final output.

This method is not mysterious — it addresses the common information blind spots that occur when a single model executes tasks.

A single model searching for information can easily miss critical sources; when handling complex constraints, it may overlook some requirements while focusing on others. Parallel exploration by multiple models allows several models with different capability structures to complement each other from various perspectives, and aggregation then reduces the impact of errors from any single model.

The OpenSquilla team presented several comparative groups in the DRACO evaluation.

In a cross-cultural e-commerce checkout process task, a single model missed the Rakuten platform's confirmation mechanism and 30-minute cancellation window — the multi-model integration solution supplemented these platform-specific details.

In a user experience analysis for AI code completion, the single model reversed the trigger strategies for debugging and feature development, while the multi-model solution delivered more accurate interaction strategies and latency budgets.

In a comparative task for Australian heat pump dryers, the single model failed to fully cover the capacity parameters of each unit, while the multi-model solution provided a more comprehensive set of comparison dimensions.

These comparisons demonstrate that single models tend to leave blind spots in complex tasks. Parallel multi-model proposals followed by aggregation can fill in some details and make results more reliable.

It is worth noting that OpenSquilla 0.5.0 did not suddenly pivot to multi-model integration. Looking back, its previous versions have been iterating around model scheduling, workflow orchestration, and result verification:

v0.1.0 implemented intelligent routing to assign simple tasks to lower-cost models; v0.3.0 introduced MetaSkill self-organizing workflows, allowing the Agent to arrange subtasks based on task objectives; v0.4.0 added verifiable coding and a signed desktop edition, putting code generation into an executable verification process. With v0.5.0, multi-model integration was brought into the Harness layer.

These successive updates have consistently pursued clear goals: making model invocation more controllable, lower-cost, and producing more stable results.

03. The Value of Domestic Models Is Shifting From Single-Point Capabilities to Combined Capabilities

The DRACO results of OpenSquilla 0.5.0 are in themselves only a phased observation. Their significance lies in providing a lens through which we can see two concurrent trends: domestic model capabilities are diversifying, and model orchestration methods are maturing.

First, let's look at the models themselves.

In the DRACO single-model evaluation, DeepSeek V4 Pro scored 50.32, Qwen 3.7 scored 49.34, and GLM-5.2 scored 48.28 — still trailing Fable5's 59.80. At the single-model level, they have not yet caught up to flagship models.

But the role of domestic models has changed. They are no longer just low-cost alternatives, and have begun to develop their own strengths across different task types. A single model may not outperform others comprehensively, but when placed in a well-designed orchestration framework, their combined value is amplified.

Now let's look at the orchestration methods.

Routing mechanisms represented by Agentic Routing have evolved from "selecting models based on questions" to "scheduling models based on execution state." In Agent scenarios, the system checks which step the task has reached, whether risks have increased, and if verification or recovery is needed — before deciding which model to invoke next.

OpenSquilla's routing design includes four layers: complexity filtering, task type classification, context state recognition, and final model ranking. The key is not any specific algorithm, but that the routing begins to understand the execution state of the task at hand.

The same task may require different models at different stages. Early decomposition can use low-cost models, switching to more appropriate models when complex evidence recovery is needed; after verification failures, more powerful models can intervene to make corrections. Routing decisions no longer only target the user's question itself, but the state of the task at its current execution stage.

Each routing decision also leaves execution data, including model selections, results, and costs. This data continues to feed back into the router and may even help train local models better suited to run within the Harness. The more accurate the routing, the more tasks can be executed under the same budget, and the greater the room for subsequent system optimization.

The performance of OpenSquilla 0.5.0 demonstrates that when the model pool is sufficiently rich and the scheduling system is sufficiently refined, complex tasks do not necessarily need to be defaulted to the most expensive model.

04. Agent Must Ultimately Focus on Delivery

The ultimate measure of AI usage lies in delivered outcomes.

Whether an Agent can integrate into real workflows depends on several basic conditions: whether the task can be completed, whether results are verifiable, and whether costs are manageable. Many past projects failed not because the model could not perform the task, but because the costs became untenable when scaled up.

The philosophy represented by OpenSquilla 0.5.0 advances cost control beyond simple downgrading to capability matching.

Simple downgrading only replaces expensive models with cheaper ones, which often leads to unstable results. Capability matching addresses which steps must invoke powerful models, which steps can be assigned to lightweight models, and which scenarios require multi-model collaboration. It seeks a balance point between cost and quality.

In this logic, cost reduction means eliminating unnecessary high-cost invocations. Complex tasks do not always need to rely on a single flagship model. As long as the model combination, execution state, and verification mechanisms are properly designed, a group of lower-cost models can achieve comparable or even better results in some tasks.

When model usage costs can be compressed through systems engineering, some projects that previously had "unfavorable ROI" gain room for re-evaluation. Users do not necessarily care how many models are invoked behind the scenes, nor do they fixate on a single model's benchmark ranking. They care more about how many tasks the system can stably complete under the same budget, whether it can recover from errors, and if the final results are deliverable.

As a result, the efficiency dividend of the AI industry is extending from the foundational model layer to the model orchestration and execution framework layer.

The 100 tasks on DRACO are only a snapshot — they cannot replace real production environments, nor long-term validation. But they remind the industry that complex tasks do not only have one path of waiting for more powerful models. Orchestrating existing models more meticulously, reliably, and affordably is also a path worthy of investment.