Author: Xiaojing, Tencent Technology
On November 7th, Microsoft released a technical white paper on its Agent Execution System, which included a set of somewhat unexpected test results: using the same Claude Opus 4.8, simply replacing the Harness behind Copilot Studio increased the success rate of complex tasks by 22.1 percentage points. The new system almost doubled the number of times the model was called for each task, but the equivalent input token consumption was reduced by 21%.
Just ten days earlier, Microsoft announced that it would roll out the unified Copilot Harness to multiple products and sell managed services to developers through its cloud platform.
On November 5th, Dark Side of the Moon released Kimi Hosted Agents, formally proposing Harness as a Service.
In August, DeepSeek open-sourced its Harness; OpenAI and Anthropic had already turned their internal operating systems into services that developers could call. The "operating system" behind the large model is beginning to become a key player in the eyes of model companies. Harness can be understood as an agent operating system built around a large model. It allows the model to use external tools, continuously execute multi-step tasks, and adjust subsequent actions based on the execution results. For agents that need to run for a long time, Harness is also responsible for saving the task state, so that work can continue after interruption. It operates at a different level from AI applications. For example, products like Claude Code and WorkBuddy are directly user-facing, while the underlying Harness can be reused by different applications or provided as an independent service to developers. As the tasks undertaken by agents become increasingly complex, Harness gradually becomes a crucial factor affecting task success rates, execution efficiency, and operating costs. Anthropic was one of the early model companies to invest in Harness development, initially to address problems encountered by Claude Code in long-running programming tasks. In late 2025, Anthropic discovered that when Claude's context window ran out, new rounds of dialogue often couldn't continue the previous work. Engineers then improved the Harness behind Claude Code, allowing the agent to save progress before the end of each stage, and then read these records in the next round to continue the unfinished task. In this way, even if a model session ends, the entire programming task can continue. In April of this year, Anthropic further launched Managed Agents, providing this agent execution capability as a cloud service to developers. To enable enterprises to run agents stably and long-term, Anthropic also adjusted the underlying architecture: Harness is responsible for driving task execution, historical records can be saved separately, and the sandbox for AI programs can be replaced at any time. Even if the sandbox malfunctions, the Agent can resume its previous work. From the internal operating system of Claude Code to a hosted service for external developers, Anthropic has transformed its years of accumulated Harness capabilities into an independent product. OpenAI has also chosen to open up the Harness behind Codex to more developers. In February of this year, OpenAI disclosed that the Codex web version, command line, and IDE plugin actually share the same Agent operating system. In August, it further positioned Codex as an open Harness platform, allowing third-party applications to directly access it. The Agents API, launched in September, is hosted and maintained by OpenAI, allowing developers to directly access the Agent's runtime environment. This approach allows OpenAI to both directly serve users through Codex and provide underlying execution capabilities to other software companies. Developers can retain their own product interface and business processes, allowing Codex to handle the specific tasks in the background. DeepSeek released the entire Harness framework code under the MIT license on August 13th, proposing "Everything is a Plugin," essentially breaking down the entire Agent runtime system into freely combinable plugins. Developers can not only replace models and external tools, but also modify how the Agent calls models, executes tasks, and even allow the Agent to write and install new plugins itself. This Cordis-based architecture aims to allow developers to develop different types of Agents according to their needs without having to build the runtime system from scratch. On November 5th, Kimi officially announced Hosted Agents, providing a 24/7 cloud environment. Enterprises can place the Agents in their internal systems or embed them into products sold externally, explicitly stating Harness as a service. Harness Evolution Harness's popularity didn't actually originate with model companies. Earlier this year, the rise of OpenClaw showed the industry that a company without a foundational model could attract a large user base with its agent-based operating system. In February, OpenAI recruited Peter Steinberger, the founder of OpenClaw, to participate in the development of the next generation of personal agents. Subsequently, Nous Research launched the Hermes Agent, attempting to allow agents to save their accumulated experience as reusable skills and continuously improve them in subsequent tasks. In late March, Stanford University and other institutions proposed Meta-Harness, which allows AI to directly modify Harness code based on failure records and verify the effects through testing. In a text classification experiment, accuracy improved by 7.7%, and context token consumption was reduced to one-quarter of the original. Why are these explorations worthwhile for model companies to invest in? In October, Nanyang Technological University in Singapore tested 66 models combined with Harness and found that changing Harness could even change the performance ranking between models. FrontierHarness Eval also found that for the same Kimi K3, the task pass rate was 66.7% using Codex, 63.3% using DeepSeek Harness Creator, and 56.7% using Kimi Code. For companies developing basic models, improving model performance relies increasingly on the design and optimization of Harness, in addition to improving the model itself. Researchers are also trying to apply the experience accumulated through Harness to model training. Co-Harness, proposed in July, first improves Harness through failure records, and then trains the model using the optimized execution data. HarnessFlow, announced in November, attempts to automatically search for better Harness designs, improving the average performance by 10.89 percentage points across six task categories. However, automatic optimization also carries risks. A study in early November found that in a specific experiment, a Harness model, after being optimized for a new task, saw its score on the SWE-bench drop from 94% to 32%. Model upgrades may also render previously effective designs ineffective. This means that the technological accumulation of Harness is unlikely to be a one-time achievement. How to maintain stability across different tasks during continuous improvement remains an unsolved problem. Industry insiders say this is also a major pain point in deploying open-source models. Therefore, model companies have begun selling Harness services. While open-source frameworks are available for free, enterprises still need to bear the deployment and maintenance costs to ensure long-term stable operation of agents. Companies like Kimi and Anthropic saw an opportunity to take over these tasks, turning Harness into an independent business, but their approaches differ significantly. Harness is starting to become a separately charged service, but 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 charged separately. Based on the official example of a one-hour programming task, the total cost is $0.58, with managed operation fees accounting for only about 14%. Kimi's pricing is divided into model tokens, sandbox operation, and paid plugins. A legal analysis task for investment banks, which it announced, was estimated to cost around 30 yuan, with sandbox fees less than 0.1 yuan. OpenAI waived additional service fees during the Agents API public beta, primarily charging based on model and tool usage. DeepSeek open-sourced Harness under the MIT license, without a direct charging logic. Therefore, the direct revenue that model companies currently obtain from Harness is likely negligible. So why invest heavily in R&D resources? A key reason is that once enterprises actually deploy Agents, their demand for Harness may be more stable than for a specific model. ABC Legal, an American legal services company, initially deployed its agents on employee computers, running them via scheduled tasks. As usage expanded, CTO Brandon Fuller began searching for a platform to centrally manage these agents. He compared several options and ultimately chose Anthropic's Managed Agents because the company didn't want to continue maintaining a fragmented operational environment. According to Anthropic, ABC Legal deployed over 50 agents to production within a month, with approximately 310 employees using Claude daily. Employees didn't even need software development knowledge to create new agents through configuration. This demonstrates that the appeal of managed agents stems from genuine enterprise needs. Rather than developing, deploying, and maintaining a system from scratch, companies may prefer to purchase off-the-shelf services, allocating engineering resources to their core business. However, companies may not be willing to become tied to a model company. In August of this year, VentureBeat released a survey of 107 companies. 85% of the surveyed companies use two or more agent orchestration platforms simultaneously. When choosing a platform, 29% of companies value the flexibility of models and tools most, while only 10% prioritize native compatibility with a leading model. Although the sample leans towards large technology companies, it reflects a purchasing trend: customers want to retain the freedom to change models. This provides Harness with another business logic for competition. If companies are willing to use a single operating system long-term, the provider, even if it cannot exclusively control model calls, may continue to receive revenue from hosting, maintenance, and enterprise development services, and establish a more stable partnership with the customer. In this way, Harness can access the actual execution process of a business's work. Compared to a simple model API, Harness can record the workflow of a task from start to finish, including model calls, tool execution, failure retries, and the final result. These records can help businesses troubleshoot problems and provide valuable feedback for model development. If validated feedback from real-world tasks can be continuously obtained, model companies have the opportunity to discover problems with their models in the production environment more quickly. However, this advantage is also subject to explicit limitations on data rights. This also raises a new question for model companies: if enterprises can freely change their underlying models while maintaining control over their own data, how much long-term advantage can Harness build? For model companies, Harness may bring more long-term value than direct fees. As enterprises delegate more and more work to agents, whoever can provide a reliable operating system will have the opportunity to influence model selection and how AI participates in real-world business. A familiar yet clichéd term: entry point. However, this term, deeply rooted in the internet age, always touches a nerve in the competition of the AI era. Now, Harness, to some extent, bears this expectation and imagination.