Scientific research tools used to each cover a separate segment of the work: ClawsGO encapsulates the entire end-to-end workflow in the cloud.
On October 26, 2026, BitMiracle AI Lab published a 23-page paper along with supporting source code: its research-oriented agent ClawsGO Science ran continuously on the cloud for 11 days, completed the formal proof of the detailed balance scenario of the global attractor conjecture proposed in 1974 in the field of chemical kinetics, produced 83 Lean proof files with approximately 9,300 lines of code, and no logical gaps were found after 21,580 compilation self-checks. This achievement has not yet been peer-reviewed, and its academic value remains to be tested by the academic community. However, as an unattended long-duration scientific research execution by an agent, it provides a verifiable technical milestone for this company established less than two years ago.
ClawsGO Science, operated by Beijing BitMiracle Technology Co., Ltd., is a cloud-based AI agent for scientific research scenarios. It has been publicly reported by media such as 36Kr since September 2026, and is currently in the early commercialization stage after public launch, pushing forward the next round of financing.
Scientific Research Processes Fragmented by Toolchains, Long-Duration Tasks Lack Stable Execution Environments
The output form of scientific research work has long been highly digitalized, but the process itself is still scattered among different tools: literature retrieval and reading, data cleaning and statistical modeling, chart drawing, paper writing and typesetting, submission format adjustment and revision. Each link has mature independent tools, but there is no connected execution link. Researchers often spend as much time on environment configuration, data migration, format alignment and version backtracking as on creative work, and these links do not constitute the core value of academic achievements.
After the popularization of general-purpose programming agents, this contradiction has not been naturally resolved. The operation cycle of general agents is usually designed around a single interaction. When facing scientific research tasks with long chains and easy failure in single steps, they are prone to memory loss or interruption during multiple jumps between retrieval, reading, modeling, drawing and writing. Scientific research tasks have two additional hard requirements: conclusions and data must be traceable to their sources, and the operating environment must be reproducible. Simply moving code generation capabilities into browsers cannot solve the long-duration nature and evidence chain problems of scientific research work.
External conditions are also converging in the same direction. In June 2026, the National Data Bureau issued the Implementation Plan for Promoting the Action of Building High-Quality Datasets in Industries, which for the first time deployed high-quality data construction in scientific research and other fields at the industry dataset level. According to calculations by market research institution QY Research, the global AI for Science market size was approximately 4.538 billion US dollars in 2025, and is expected to reach 26.23 billion US dollars in 2032, with a compound annual growth rate of about 28.9% during the period, among which software platforms are one of the most important product forms. The convergence of demand, policies and declining computing power costs at the same time window has made scientific research agents an independent product layer beyond the capabilities of foundational models.
Encapsulate Environment, Execution and Delivery into a Cloud-Based Scientific Research Machine
The product form of ClawsGO Science can be summarized as: encapsulating the environment, tools and processes that originally required researchers to manually string together locally into a scientific research machine running on the cloud. The product allocates a cloud host pre-installed with a scientific research environment for each task. Users describe the research goals in natural language, and the agent independently breaks them down into subtasks and pushes forward continuously. The task status is displayed synchronously on the web page, mobile terminal and desktop client. When local data or self-owned computing power needs to be accessed, users can also switch the task to the local machine, or access the laboratory GPU host via SSH. The product provides desktop clients for both macOS and Windows.
At the architecture level, the product adds two layers of engineering capabilities on top of the general agent kernel: the first is the long-duration execution mechanism, where tasks can run continuously for several hours around the research goal and resume from the breakpoint after interruption; the second is the task graph and multi-sub-agent collaboration mechanism, where complex tasks are split into subtasks that can be pushed forward in parallel, and users can view the overall progress and status of each subtask on the panel. The writing section provides a native LaTeX workspace, where source code and PDF are compiled in real time for comparison; scientific research drawings are delivered in the form of reproducible packages, and each drawing retains the corresponding code, data and operating environment records. All traces of environment, commands, file modifications and retrieval actions of each execution are retained, and the whole round of work can be replayed at any time. This design directly responds to the requirements of the scientific research scenario for the evidence chain.
The formal verification of the global attractor conjecture published in October is a concentrated test of the above architecture. According to the team's public papers and reports from media such as The Paper and CSDN, this task adopts a two-layer multi-agent architecture where the decision-making model grasps the direction and more than 100 execution models derive in parallel, and the mathematical theorem prover Lean undertakes deterministic verification. After the commonly used textbook route was blocked on the 6th day, the decision-making model switched to the elementary geometric energy descent route, and then the full compilation passed. The academic recognition of such achievements is still pending peer review, but its engineering implication is already clear: the agent can work continuously for more than a week around a serious scientific problem under unattended conditions, and deliver machine-verifiable conclusions.
Enter Individual Researcher Market via Subscription and Computing Power Billing, Plan Private Deployment on the Institutional Side
In terms of business model, the product adopts membership subscription plus computing power billing as the main charging method: AI usage is converted into computing power according to tokens, and the model actually used and consumption of each step can be checked one by one; the server running time is not billed separately, and the computing power cost is integrated into the model calling service. This structure limits the variable cost of researchers to the task itself, does not require users to retain idle computing power for a long time, and avoids the one-time investment of local high-performance equipment. Target users cover graduate students, university teachers, independent researchers and enterprise R&D teams. The former's demand is to shorten the daily scientific research process, while the latter pay more attention to process specification, data isolation and private deployment; for institutional customers, the team plans to provide private deployment and pre-positioned engineer co-construction mode. At this stage, the product is still in the parallel stage of product verification and early commercialization, and the specific paying user scale, repurchase rate and revenue data have not been publicly disclosed.
From Campus Tools to Scientific Research Agents, Narrow the Battlefield Around Emergent Demands
Founder Lü Haoran was born in 2004, is a computer science student at Jilin University, suspended his studies in the sophomore year to start a business. He previously worked as an intern AI product manager at Baidu for about four months, and then interned at the venture capital institution Qichuang Chuangtan for about two months. The starting point of his entrepreneurship was AI account tools on campus and an auxiliary product for mathematical modeling competitions in early 2025; at the end of 2025, the company tried to cover general agents for various professional scenarios, and then found that about half of the early users were doing scientific research, many of whom were teachers and researchers. The team then narrowed the product to scientific research scenarios in 2026, forming the current ClawsGO Science.
The milestones at the company level are equally clear: the operating entity Beijing BitMiracle Technology Co., Ltd. was established in November 2024, with a registered capital of 100,000 RMB, and the legal representative is Lü Haoran. Its industrial and commercial registration business scope covers software development and artificial intelligence application software development; the team also set up BitMiracle (Changsha) Technology Co., Ltd., which was listed on the Hunan Equity Exchange College Student Innovation and Entrepreneurship Special Board in August 2026, with the listing code 310293HN. The product has launched two independent entrances for the China version and the global version. According to previous reports from 36Kr, the funds for the next round of financing are planned to be used in three aspects: engineering optimization of long-duration task scheduling and multi-model routing, productization of enterprise-level private deployment, and expansion of the scientific research team.
Competition in the track is extending from general programming assistants to dedicated scenarios. Both foundational model manufacturers are extending to the scientific research field, and vertical startups are cutting in from experimental design and literature mining. The relative position of ClawsGO Science is to take the entire scientific research process as the product boundary, rather than only making single-point functions; the formal verification achievement in October provides a verifiable technical footnote for this route. As for the success rate of long-duration tasks, interdisciplinary universality and paying team retention data, they are the answers that this company needs to continue to provide to the market in the next financing cycle.