Text | Xiaojing

Editor | Xu Qingyang

On November 7, Microsoft released a technical white paper on the Agent execution system, containing a set of somewhat surprising test results: using Claude Opus 4.8, simply replacing the Harness behind Copilot Studio improved the pass rate for complex tasks by 22.1 percentage points. The new system nearly doubled the number of times a model was invoked for each task, yet the adjusted input Token consumption could be reduced by 21%.

Just ten days prior, Microsoft had announced the promotion of a unified Copilot Harness across several of its products, offering managed services to developers through its cloud platform.

On November 5, the dark side of the moon released Kimi Hosted Agents (managed agents), officially proposing Harness as a Service. In August, DeepSeek open-sourced its own Harness; OpenAI and Anthropic had already transformed their internal operating systems into services callable by developers.

The "operating system" behind that large model has begun to emerge as a key player in the eyes of model companies.

Image generated by AI

01 Seizing the Harness

The Harness can be understood as a set of Agent operating systems built around large models. It enables models to use external tools, continuously execute multi-step tasks, and adjust subsequent actions based on execution results. For Agents that require long-term operation, the Harness is also responsible for saving task states, allowing work to continue after interruptions.

It exists at a different level than AI applications. For example, products like Claude Code and WorkBuddy are directly aimed at users, while the underlying Harness can be reused by various applications and offered as an independent service to developers. As the work undertaken by Agents becomes increasingly complex, the Harness has gradually become an important factor influencing task success rates, execution efficiency, and operational costs.

Anthropic is one of the earlier model companies to invest in Harness development, initially to address issues Claude Code faced during prolonged programming tasks.

By the end of 2025, Anthropic discovered that when Claude's context window ran out, subsequent dialogues often could not continue previous work. Engineers then improved the Harness behind Claude Code, allowing the Agent to save progress at the end of each stage and read these records in the next round to continue unfinished tasks. In this way, even if a model session ends, the entire programming task can still be advanced.

In April this year, Anthropic further launched Managed Agents, offering this set of Agent operating capabilities as a cloud service for developers. To enable enterprises to run Agents long-term and stably, Anthropic also adjusted the underlying architecture: Harness drives task execution, historical records can be stored separately, and the sandbox for AI programs can be replaced at any time. Even if the sandbox fails, the Agent can revert to previous work.

From the internal operating system of Claude Code to hosted services for external developers, Anthropic has turned years of accumulated Harness capabilities into independent products.

OpenAI also chose to open the Harness behind Codex to more developers. In February of this year, OpenAI revealed that the Codex web version, command line, and IDE plugins actually share the same set of Agent operating systems. In August, it further positioned Codex as an open Harness platform, allowing third-party applications to directly connect. The Agents API launched in September, hosted and maintained by OpenAI, enables developers to directly access the Agent's operating environment.

This route allows OpenAI to serve users directly through Codex while providing underlying execution capabilities to other software companies. Developers can retain their product interfaces and business processes, letting Codex handle specific tasks in the background.

On August 13, DeepSeek open-sourced the entire framework code for Harness under the MIT license, proposing "Everything is a Plugin," almost breaking the entire Agent operating system into freely combinable plugins. Developers can not only replace models and external tools but also modify how Agents invoke models and execute tasks, and even enable Agents to write and install new plugins themselves.

This Cordis-based architecture aims to allow developers to develop different types of Agents according to their needs without having to build an operating system from scratch.

On November 5, Kimi announced that Hosted Agents would provide a 24/7 cloud environment, allowing enterprises to place Agents into internal systems or embed them in products sold externally, explicitly proposing Harness as a service.

02 Evolution of Harness

In fact, the popularity of Harness did not start with model companies.

Earlier this year, the rise of OpenClaw already showed the industry that a company without a foundational model could attract a large number of users through an Agent operating system. In February, OpenAI recruited OpenClaw founder Peter Steinberger to participate in the development of the next generation of personal Agents.

Following this, Nous Research launched the Hermes Agent, attempting to have Agents save accumulated experiences from work as reusable Skills and continually improve them in subsequent tasks.

At the end of March, institutions like Stanford University proposed Meta-Harness, which allows AI to directly modify Harness code based on failure records and verify effectiveness through testing. In a text classification experiment, accuracy improved by 7.7%, and context Token consumption was reduced to a quarter of the original.

These explorations raise the question: why is this worth investing in by model companies? In October, Nanyang Technological University in Singapore tested combinations of 66 models with Harness, finding that changing Harness could even alter performance rankings between models.

FrontierHarness Eval also found that for the same Kimi K3, the task pass rate using Codex was 66.7%, using DeepSeek Harness Creator was 63.3%, and using Kimi Code was 56.7%.

For foundational model companies, enhancing model performance now increasingly relies on the design and optimization of Harness, in addition to improving the models themselves.

Researchers are also trying to apply the experience accumulated by Harness to model training. The Co-Harness proposed in July aims to first improve the Harness based on failure records and then use the optimized execution data to train models. The HarnessFlow announced in November attempts to automatically search for better Harness designs, improving average scores by 10.89 percentage points across six task types.

However, automatic optimization also carries risks. A study from early November found that in specific experiments, a Harness optimized for new tasks saw its original score on SWE-bench drop from 94% to 32%. Model upgrades may also render previously effective designs ineffective.

This means that the technical accumulation of Harness is difficult to achieve in a one-time effort. How to maintain the stability of different tasks during continuous improvement remains an unresolved challenge.

Industry insiders have indicated that this is also a major pain point for deploying open-source models. Therefore, model companies have begun selling Harness services. Open-source frameworks can be obtained for free, but if enterprises want to run Agents stably and long-term, they still need to bear the costs of deployment and maintenance. Companies like Kimi and Anthropic have seen opportunities in taking on this work, turning Harness into a standalone business, but different companies have significantly different approaches.

03 The Hidden Value of Harness

Harness is beginning to emerge as a separately chargeable service, but the current pricing shows that model inference remains the primary source of revenue.

Anthropic's Managed Agents charge $0.08 per active session hour, with Tokens billed separately. According to an official example of a one-hour programming task, the total cost is $0.58, with only about 14% attributed to managed operation fees.

Kimi's pricing is divided into model Tokens, sandbox operations, and paid plugins. It estimates the cost of a legal analysis task for an investment bank to be about 30 yuan, with sandbox fees under 0.1 yuan. OpenAI waived extra service fees during the public testing period of the Agents API, billing primarily based on model and tool usage. DeepSeek has open-sourced Harness under the MIT license, with no direct charging logic.

Therefore, it can be said that the direct revenue that model companies currently obtain from Harness may be negligible.

So, why still invest significant R&D resources?

An important reason is that once enterprises truly deploy Agents, the demand for Harness may be more stable than for any specific model.

The American legal services company ABC Legal originally deployed Agents on employee computers, running them through scheduled tasks. As the scope of use expanded, CTO Brandon Fuller began looking for a platform that could manage these Agents uniformly. He compared several options and ultimately chose Anthropic's Managed Agents because the company did not want to continue maintaining a dispersed operating environment.

According to Anthropic, ABC Legal quickly put more than 50 Agents into production within a month, with approximately 310 employees regularly using Claude. Employees didn't even need to have software development knowledge to create new Agents through configuration.

This indicates that the appeal of a managed Harness arises from actual corporate needs. Compared to building and maintaining operational systems from scratch, enterprises may prefer to purchase ready-made services, redirecting engineering resources toward their own business.

However, enterprises may not wish to bind themselves to a single model company.

In August this year, VentureBeat published a survey of 107 companies. 85% of participating companies use two or more Agent orchestration platforms simultaneously. When choosing platforms, 29% of companies prioritize model and tool flexibility, while only 10% prioritize native compatibility with a leading model. Although the sample leans toward large technology companies, it already reflects a purchasing tendency: customers wish to retain the freedom to switch models.

This provides another commercial logic for Harness competition. If enterprises are willing to use a single operating system long-term, providers, even if they cannot monopolize model calls, might still gain ongoing income from hosting, maintenance, and enterprise development services, thereby establishing more stable cooperative relationships with clients.

Thus, Harness can engage with the actual execution processes of enterprises' work. Compared to purely model APIs, Harness can document the work trajectory of a task from start to finish, including model calls, tool executions, retry failures, and final results. These records can help enterprises troubleshoot failures and provide valuable feedback for model development.

If real task feedback can be continuously obtained, model companies will have the opportunity to discover the issues of models in production environments faster. However, this advantage will also be subject to clear limitations regarding data rights.

This also poses a new question for model companies: if enterprises can freely switch underlying models while insisting on retaining their data, how much long-term advantage can Harness still establish?

For model companies, Harness may bring more long-term value than direct charges. As enterprises delegate more work to Agents, those who can provide reliable operating systems will have the opportunity to influence model selections and how AI is integrated into real business.

A familiar yet clichéd term: gateway.

However, this term, deeply rooted in the Internet era, will always touch the sensitive nerves of competition in the AI era. Now, Harness has, to some extent, taken on this expectation, as well as imagination.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到[email protected]，本平台相关工作人员将会进行核查。