Recently, CREAO AI, an early-stage portfolio company of HyT Capital, completed a new strategic financing round of several million dollars. CREAO AI is exploring the creation of self-evolving infrastructure for enterprises (Agent OS), accelerating, optimizing, and even automating the process of Agent building by enabling AI to continuously improve itself. 

This article is republished with permission from "AIEmergence"; Author: Wang Xinyi; Editor: Zhang Yuxin.

RSI (Recursive Self-Improvement) is becoming one of the hottest concepts in the AI space this year.

RSI, or Recursive Self-Improvement, is fundamentally about using AI to train AI and enable self-improvement. RSI can include three levels: Artifacts — self-evaluating and optimizing the outputs produced; Harness — optimizing the prompts, memory, tools, skills, and other elements used by Agents when executing tasks; and Models — autonomously optimizing model parameters or training the next generation of models.

Last month, Weng Li, co-founder of Thinking Machines Lab, officially announced her departure and returned to OpenAI to lead the RSI team. Earlier, she published a widely read blog post titled "Harness Engineering for Self-Improvement," in which she suggested that RSI is likely to first break out through Harness Engineering.

CREAO has quickly turned its focus toward RSI in the Harness layer. For the B2B market, CREAO provides cloud-based infrastructure services that allow enterprise clients to create, run, and manage their own Agents on the platform. Through a simple conversation, users can generate and deploy a fully functional, reusable application. In this process, AI can generate tools, execute autonomously, and use tools, while humans only need to define requirements and handle approvals.

Today, CREAO is focusing on using AI for self-improvement — accelerating, optimizing, and even automating the Agent-building process. The problem CREAO aims to solve is whether AI, in real-world scenarios, can accumulate experience from its own execution, feedback, and corrections, and then apply that back to its own self-improvement, enabling better performance the next time it executes a task.

It is important to clarify that existing Agents already have the ability to remember a user's context, historical tasks, and preferences, making the experience increasingly smooth for the same user over time. However, this does not raise the upper limit of the Agent's capabilities. Self-improvement, on the other hand, refers to the Agent's ability to iterate its own model weights or Harness based on the execution and feedback from a large number of real-world tasks, making itself more capable.

Currently, CREAO is working on both: the personalized memory capability on one hand, and the exploration of self-improvement on the other. At this stage, leading labs both domestically and internationally have not yet achieved significant results in model self-improvement. What CREAO has already validated is limited, data-driven self-iteration of the Harness within a specific business scenario.

AIEmergence has exclusively learned that CREAO has recently completed a new strategic financing round of several million dollars. The funds will be primarily used to expand real-business workflows and the scale of high-quality feedback data, accelerate the iteration of vertical models through post-training based on open-source models, and further improve the runtime environment for Agent self-evolution. Previously, CREAO had accumulated over $30 million in total financing.

CREAO's core team members come from companies including Meta (Llama team), Apple, LinkedIn, TikTok, MiniMax, Alibaba, and ByteDance. CREAO's current client base is primarily concentrated among SMEs, with main application scenarios in digital marketing, e-commerce, and real estate. After analyzing retained users, 65.2% of tasks come from business and finance-related scenarios.

Following the launch of its cloud-based services, CREAO is also exploring better ways to deploy locally within enterprise environments. The company is set to release a new version of its product tailored for enterprise scenarios, addressing customer needs around data security, compliance requirements, and private deployment options.

Cheng Kai, Co-founder and CEO of CREAO AI, believes that "the battle in the B2B enterprise market hasn't really started yet." Starting with SMB customers is not the endgame — CREAO is gradually shifting its market focus toward larger clients, with the goal of extending and deploying enterprise Agent self-improvement capabilities in higher-volume scenarios.

01

Self-Evolution at the Harness Layer Offers Higher ROI Than Foundation Models

Over the past six months, discussions around Harness have not stopped. Some view Harness as a way to enhance model intelligence. Cheng Kai offers a different perspective: Harness is essentially an environment layer, and its more important value lies in solving the feasibility problem of model application.

To use an analogy: if you put a PhD graduate in an empty room with only paper and pen, no matter how intelligent they are, they cannot produce high-quality output. But if you put a middle school student in a room with internet access and various tools, what they can achieve may surpass that of the PhD.

Cheng Kai believes that Harness is that room and the tools within it. Without a well-designed Harness architecture and capabilities, even the best large models and Agents will struggle to perform effectively in real application scenarios.

It would be even better if this "room" could continuously improve itself. On the topic of RSI, CREAO has taken the lead in exploring self-evolution at the Harness layer. Specifically, this means achieving self-improvement of the Harness through real-world production tasks and feedback mechanisms: AI operating in real production environments accumulates experience and self-corrects through every execution, feedback, and correction, turning those experiences into better performance the next time around.

CREAO has access to real users, industry workflows, and evaluations from professional users — all of which become the data that drives Harness self-improvement. After each round of real-world tasks, the system can use this data to optimize Agent planning, tool invocation, memory, task scripts, model routing, and fault recovery mechanisms, enabling direct improvements at both the Agent and Harness layers.

Cheng Kai told us: "Compared to fine-tuning foundation models, self-iteration at the Harness layer offers higher ROI. Solving operational issues is far faster than solving model intelligence problems."

However, working on the Harness alone is not enough. AI self-iteration and self-evolution consist of two levels: first, self-improvement based on errors made when connecting to databases or reading documents — these are objective errors; Second, self-improvement based on the quality of task execution results — these are subjective errors.

Self-evolution at the Harness layer can only address objective errors. To correct subjective errors and achieve true self-evolution, exploration at the foundation model layer is also required. For this reason, CREAO is now launching foundation models that undergo continuous post-training based on business data.

Many frontier labs are conducting research on foundation model self-evolution. CREAO does not intend to follow the narrative of pure technology exploration — its goal is not to train an increasingly intelligent model that can beat all others. As open-source models continue to improve in quality, they see a faster path to model self-iteration through commercialization in vertical scenarios and access to high-quality scenario data.

So what they are doing now is data-driven self-iteration — continuously training models that can learn, validate, and improve in real production environments. These models aim to achieve two things: first, lower user costs by reducing token consumption and making tokens more efficient; second, train models driven by actual operational metrics — using more efficient, lower-cost models in specific scenarios to achieve task performance close to that of frontier models.

In the coming months, CREAO will be the first to launch a foundation model for a core scenario, based on post-training of open-source models. They will also continue to release more self-iterating models for key scenarios on the platform, including e-commerce, real estate, logistics and transportation, and digital marketing. On this foundation, users will be able to use more efficient, lower-cost models to achieve task performance close to that of frontier models.

"Many enterprises simply cannot afford the most advanced models," said Cheng Kai. "The core purpose of our post-training efforts is also to reduce costs for users."

02

The Core of Self-Evolution Is the Quality of Real-World Scenario Data

In just over four months since its launch, CREAO AI has accumulated over 270,000 Agents created on the platform. Since the platform primarily serves SMBs and professional users, it has naturally amassed a large volume of real vertical scenario data and high-quality user feedback. Cheng Kai believes that data quality is critical — without high-quality data, even strong technical self-improvement may not deliver the desired results.

The value of this data lies in its ability to address three core challenges in AI product commercialization: how to solve tasks better, faster, and with greater stability and efficiency. CREAO offers solutions across all three dimensions.

First, the outcome of model and Harness self-evolution. Based on the data accumulated from real business scenarios and user feedback, CREAO can integrate post-training of its models with Harness iteration to form a production system capable of self-evolution. The system can then determine when to call the most advanced frontier models, when to use its own vertical models, which insights should remain at the Agent layer, and which should be trained into model parameters.

The most immediate benefit of this approach is a significant reduction in cost and improvement in efficiency. According to its internal benchmark tests, the fine-tuned models deployed in the CREAO system are faster and cheaper to run than calling external APIs. Cheng Kai told us that an increasing number of tasks on the CREAO platform are now running on open-source models, delivering the same task quality to users while saving both cost and time.

Second, stability. Cheng Kai believes that during the GPT-3.5 era, many tasks could already be automated, but commercial deployment was still unattainable because models at the time could not guarantee stability and security.

They have observed that over 95% of Agents on the CREAO platform are run daily, with the same tasks potentially executed hundreds or thousands of times. At that scale, platform stability and error reduction become critically important from a commercialization standpoint. CREAO's approach is to strengthen Harness self-repair and self-optimization, using accumulated data from similar tasks to refine Harness capabilities.

E-commerce is a classic example. In e-commerce scenarios, many users make similar requests, so the platform accumulates a large volume of similar workflows. For instance, when using Agents for SEO optimization — although each person's approach may vary slightly — the general direction is largely the same. In this context, the platform can aggregate user feedback data at a macro level, improve overall Harness self-improvement capabilities, and even optimize Agent capabilities across all users at the platform level.

The result of Harness self-improvement is that when new users want to build an SEO Agent, they no longer need to go through repeated fine-tuning to achieve satisfactory results, unlike the first cohort of users.

Third, long-term high-quality business data. Cheng Kai told us that among retained users, 93.4% of operations are triggered automatically by the system. Furthermore, over 60% of automated workflow tasks have been running continuously for more than a month. This means AI has genuinely entered users' work scenarios, naturally providing CREAO with sustained, comparable business execution data.

It is worth noting that, much like its products, the CREAO team also carries a strong AI-First ethos. Peter, Co-founder and CTO of CREAO, posted on X about a thread that accumulated nearly 2 million views, mentioning that the team's product development speed is exceptionally fast — 99% of the code is written by AI, and they can achieve what traditionally takes six weeks of development in just one day.

In Cheng Kai's view, this AI-First approach is not just about development methodology — it is also where the product's endgame lies. "The users of future productivity tools will not be humans, but AI. Only when AI becomes the user of productivity tools can the productivity efficiency of society or enterprises truly increase exponentially." What CREAO wants to build is not just a self-evolving product, but self-evolving infrastructure for enterprises — an Agent OS — where enterprise Agents can run, be managed, and iterate repeatedly on this operating system.