The Harness Wars: Why Model Companies Are Racing to Turn Agent Runtime Systems into Services

On November 7, Microsoft released a technical white paper on Agent execution systems, containing a set of somewhat unexpected test results: using the same Claude Opus 4.8, merely swapping out the Harness behind Copilot Studio raised the pass rate on complex tasks by 22.1 percentage points. The new system calls the model nearly twice as many times per task, yet the converted input Token consumption can be reduced by 21%.

Just ten days earlier, Microsoft had announced that it would roll out a unified Copilot Harness across several of its products and sell hosted services to developers through its cloud platform.

On November 5, Moonshot AI had just released Kimi Hosted Agents, formally proposing Harness as a Service. In August, DeepSeek open-sourced its own Harness; OpenAI and Anthropic had already turned their internal runtime systems into services developers can call even earlier.

The "runtime system" behind the large model is starting to become the key player in the eyes of model companies.

01 Racing to Secure a Foothold in Harness

Harness can be understood as a set of Agent runtime systems built around a large model. It enables the model to use external tools, continuously execute multi-step tasks, and adjust subsequent actions based on execution results. For Agents that need to run for long periods, the Harness is also responsible for saving task state so that work can continue after an interruption.

It sits at a different layer from AI applications. For example, products such as Claude Code and WorkBuddy directly face users, while the underlying Harness can be reused by different applications and can also be provided to developers as an independent service. As the work Agents take on becomes increasingly complex, the Harness is gradually becoming an important factor affecting task success rates, execution efficiency, and operating costs.

Anthropic is one of the model companies that invested in Harness R&D earlier, initially to solve the problems Claude Code encountered in long-running programming tasks.

At the end of 2025, Anthropic found that when Claude's context window was exhausted, a new round of conversation often could not continue the previous work. Engineers then improved the Harness behind Claude Code so that the Agent saves its progress before the end of each stage, and the next round reads these records to continue completing unfinished tasks. In this way, even if a model session ends, the entire programming task can still keep moving forward.

In April this year, Anthropic further launched Managed Agents, providing this Agent runtime capability to developers as a cloud service. To enable enterprises to run Agents stably over the long term, Anthropic also adjusted the underlying architecture: the Harness is responsible for driving task execution, history records can be stored separately, and the sandbox running the AI program can be replaced at any time. Even if the sandbox fails, the Agent can recover its previous work.

From the runtime system inside Claude Code to a hosted service for external developers, Anthropic has turned its years of accumulated harness capabilities into an independent product.

OpenAI also chose to open the Harness behind Codex to more developers. In February this year, OpenAI disclosed that the Codex web version, command line, and IDE plugin actually share the same Agent runtime system. In August, it further positioned Codex as an open Harness platform, allowing third-party applications to connect directly. The Agents API launched in September has OpenAI responsible for hosting and maintaining this system, so developers can directly access the Agent's runtime environment.

This route allows OpenAI both to serve users directly through Codex and to provide the underlying execution capability to other software companies. Developers can keep their own product interfaces and business processes while letting Codex complete specific tasks behind the scenes.

On August 13, DeepSeek open-sourced the entire Harness framework code under the MIT license. It proposed "Everything is a Plugin," breaking almost the entire Agent runtime system into freely composable plugins. In addition to replacing models and external tools, developers can also modify how the Agent calls models and executes tasks, and even let the Agent write and install new plugins itself.

This architecture based on Cordis aims to let developers develop different types of Agents according to their own needs without having to build a runtime system from scratch.

On November 5, Kimi officially announced that Hosted Agents provide an around-the-clock cloud environment; enterprises can put Agents into internal systems or embed them in products sold externally, and it explicitly proposed Harness as a service.

02 The Evolution of Harness

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

Earlier this year, the rise of OpenClaw had already shown the industry that a company without a foundation model could also attract a large number of users through an Agent runtime system. In February, OpenAI recruited OpenClaw founder Peter Steinberger to participate in the development of the next-generation personal Agent.

Subsequently, Nous Research launched Hermes Agent, attempting to let the Agent save the experience accumulated during work as reusable Skills and continuously improve them in subsequent tasks.

At the end of March, Stanford University and other institutions proposed Meta-Harness, letting AI directly modify Harness code based on failure records and verify the results through testing. In a text classification experiment, accuracy increased by 7.7%, and context Token consumption dropped to one quarter of the original.

Why are these explorations worth the investment of model companies? In October, Nanyang Technological University in Singapore tested 66 combinations of models and Harness and found that changing the Harness may even change the performance ranking among models.

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

For foundation model companies, improving model performance increasingly depends not only on improving the model itself, but also on the design and optimization of the Harness.

Researchers are also trying to use the experience accumulated by Harness for model training. Co-Harness, proposed in July, first improves the Harness through failure records and then uses the optimized execution data to train the model. HarnessFlow, announced in November, attempts to automatically search for better Harness designs and raised the average score by 10.89 percentage points across six types of tasks.

However, automatic optimization also carries risks. A study in early November found that in a specific experiment, after a Harness was optimized for a new task, its previous score on SWE-bench fell 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 once and for all. How to maintain stability across different tasks amid continuous improvement remains an unsolved problem.

Some industry insiders say this is also a major pain point in deploying open-source models. Open-source frameworks can be obtained for free, but if enterprises want to run Agents stably over the long term, they still need to bear deployment and maintenance costs. Companies such as Kimi and Anthropic have seen the opportunity to take on this work and turn Harness into an independent business, but different companies take very different approaches.

03 The Hidden Value of Harness

Harness has begun to become a service that can be charged for separately, but the current price list shows that model inference is still the main source of revenue.

Anthropic's Managed Agents charges $0.08 per active session hour, with Tokens billed separately. According to the official example of a one-hour coding task, the total cost is $0.58, of which the managed runtime fee accounts for only about 14%.

Kimi's charges are divided into model Tokens, sandbox runtime, and paid plugins. For a published investment banking legal analysis task, the estimated cost is about 30 yuan, with the sandbox fee less than 0.1 yuan. OpenAI waives additional service fees during the public beta of the Agents API and mainly charges based on model and tool usage. DeepSeek open-sourced Harness under the MIT license and has no direct charging logic.

Therefore, the direct revenue that model companies may obtain from Harness at this stage is likely to be minimal.

So why still invest substantial R&D resources?

An important reason is that after enterprises actually deploy Agents, their demand for Harness may be more stable than for any particular model.

U.S. legal services company ABC Legal originally deployed Agents on employee computers and ran them through scheduled tasks. As usage expanded, CTO Brandon Fuller began looking for a platform that could uniformly manage these Agents. He compared multiple options and ultimately chose Anthropic's Managed Agents because the company did not want to continue maintaining scattered runtime environments.

According to Anthropic, ABC Legal put more than 50 Agents into production within a month, and about 310 employees use Claude daily. Employees do not even need software development knowledge to create new Agents through configuration.

This shows that the appeal of hosted Harness comes from real enterprise demand. Rather than developing deployment and long-term runtime maintenance systems from scratch, enterprises may be more willing to buy ready-made services and direct their engineering resources toward their own business.

But enterprises are not necessarily willing to become tied to one model company because of this.

In August this year, VentureBeat published a survey of 107 enterprises. 85% of the surveyed enterprises use two or more Agent orchestration platforms at the same time. When choosing a platform, 29% of enterprises value the flexibility of models and tools the most, and only 10% prioritize native compatibility with a leading model. Although the sample skews toward large technology enterprises, it already reflects a procurement tendency: customers want to retain the freedom to switch models.

This precisely provides another business logic for Harness competition. If enterprises are willing to use one runtime system for the long term, even if the provider cannot monopolize model calls, it may still continuously earn revenue from hosting, maintenance, and enterprise development services, and build a more stable cooperative relationship with customers.

In this way, Harness can gain access to the execution process of enterprises' actual work. Compared with a simple model API, Harness can record the work trajectory of a task from start to finish, including model calls, tool execution, failure retries, and final results. These records can not only help enterprises troubleshoot failures, but also provide valuable feedback for model R&D.

If verified feedback from real tasks can be continuously obtained, model companies will have the opportunity to discover problems with models in production environments more quickly. However, this advantage will also be clearly limited by data rights.

This also raises a new question for model companies: if enterprises can freely switch underlying models and insist on controlling their own data, how much long-term advantage can Harness establish?

For model companies, Harness may bring longer-term value than direct charging. As enterprises hand more and more work to Agents, whoever can provide a reliable runtime system will have the opportunity to influence model choices and how AI participates in real business.

A familiar and clichéd word: gateway.

However, this term, deeply rooted in the internet era, always touches a sensitive nerve in AI-era competition. Now, Harness has to some extent taken on this expectation, and imagination.

This article comes from the WeChat public account "Tencent Technology," author: Worth Watching, published by 36Kr with authorization.