In-Depth Review of ScienceClaw: A New Paradigm for Scientific Agents from "Task Execution" to "Project Management"

1. What is ScienceClaw

ScienceClaw is a scientific native agent platform launched by ZKenny (incubated by the Institute of Automation, Chinese Academy of Sciences). It is built upon the ZKenny 4.0 multimodal large model and covers fields such as life sciences, materials science, chemistry, physics, and astronomy. The latest upgraded AutoProject engine has shifted AI from executing individual scientific tasks to autonomously managing complete research projects. It can independently perform project planning, task decomposition, long-term execution, evidence verification, and fault tolerance repair, ultimately generating reusable scientific assets such as papers, data, models, and code, while supporting human-in-the-loop collaboration. The core value of this platform lies in elevating scientific research from "task-level automation" to "project-level autonomy," providing researchers with an intelligent assistant that supports the entire research lifecycle.

Image source: Official article
Image source: official article

Technical positioning and domain: ScienceClaw belongs to the intersection of artificial intelligence and scientific discovery, specifically positioned as a scientific native agent platform. Unlike general-purpose AI assistants, it is deeply customized for scientific research scenarios, integrating the reasoning capabilities of large models with the standardization and traceability requirements of scientific workflows, offering end-to-end support across multiple disciplines including life sciences, materials science, chemistry, physics, and astronomy.

Development background: This platform was developed by the ZKenny team, which was incubated by the Institute of Automation, Chinese Academy of Sciences, and the Wuhan Artificial Intelligence Research Institute. The team has accumulated a solid technical foundation on the ZKenny multimodal large model series. The motivation for development stems from common issues in scientific research scenarios, such as labor-intensive repetitive tasks, long experimental iteration cycles, and difficulties in reusing research outcomes. The goal is to free researchers from tedious execution processes through AI agents.

Core value: ScienceClaw addresses the core pain points of traditional AI-assisted scientific research tools, which can only perform single-point tasks and are unable to manage full research topics. It achieves a transformation from "tool" to "collaborator," capable of independently managing an entire scientific research project. Furthermore, through its self-developed EviGraph evidence graph system, every scientific conclusion is supported by experiments and data, achieving a 40.74% improvement in traceable evidence rate compared to the optimal baseline, thus balancing efficiency and credibility.

Technical features: The platform adopts a hierarchical autonomous multi-agent architecture, dynamically scheduling specialized agents such as subject matter, code, search, and simulation to collaborate in a network. It uses the ZKenny 4.0 multimodal large model as its foundation, enabling unified understanding of heterogeneous scientific information such as formulas, charts, experimental curves, code, and data in research papers. Through the Project2Task planning mechanism and the TaskExecutor long-range execution mechanism, it adapts to the real nonlinear iterative nature of scientific research work.

2. Key Features

• Project2Task (Project-level Planning): Input vague scientific ideas, and the system automatically breaks them down into a structured task network and plans the research path. This mechanism integrates research objectives, literature evidence, resource constraints, and task dependencies for global modeling, selects the optimal decomposition strategy through signal/noise analysis, supports four project topologies: horizontal, vertical, horizontal-first-then-vertical, and vertical-first-then-horizontal, and plans the serial/parallel relationships of tasks and cross-task asset reuse.

• TaskExecutor (Long-term Autonomous Execution): Breaks through the linear interaction model of "one call, one return," building an autonomous research cycle driven by objectives. AI can remain on-site for extended periods to advance the project, globally monitor the project status, and continuously capture experimental feedback. When a task fails or results are abnormal, the system automatically backtracks for evaluation, rebuilds the task network, and reruns the experiments, supporting research exploration cycles that can last for months.

• EviGraph (Evidence-driven Verification): Organizes the research process into a dynamic evidence chain of "Research Question → Research Gap → Hypothesis → Experiment → Discovery → Conclusion," operating in three phases: initial graph construction, inspection and iterative repair, and manuscript generation. The system continuously verifies logical consistency across tasks, identifies root causes along the evidence chain when deviations are detected, and autonomously fixes them, with support for short-term and long-term graph libraries to enable version rollback and experience accumulation.

• Literature Search and Review: Relying on a self-built library of hundreds of millions of interdisciplinary academic papers, the system autonomously completes literature research and key finding extraction. Researchers no longer need to manually search and read through massive papers; the system can complete what used to take hours of literature research and summary work in just a few minutes, providing evidence support for research planning.

• Data Analysis and Code Execution: Uniformly schedules professional tools such as code, data analysis, and simulation to complete scientific computing tasks. Whether it's data processing, model training, or simulation, the system can automatically select the appropriate toolchain and execute it, covering the complete computational workflow from data preprocessing to result visualization.

• Scientific Asset Accumulation: After project completion, the system automatically outputs papers, datasets, models, code, and full experimental records, forming a reusable and traceable scientific asset matrix. This mechanism eliminates the drawbacks of traditional research that prioritizes papers over assets, ensuring that every research investment becomes a long-term, iterable asset.

• Human-Machine Collaboration: Researchers can intervene at any node during project progression to evaluate and correct research paths, hypotheses, and intermediate results. The system supports flexible switching between "fully automated execution" and "manual review" modes, ensuring efficiency while preserving researchers' final control over the research direction.

• Multi-platform Access: In addition to the web interface, the system also supports calling research tools via WeChat and Feishu plugins. Researchers can initiate research tasks, check project progress, and retrieve results directly within their commonly used office and collaboration environments, reducing the usage barriers of the tools.

3. How to Use

1. Environment Requirements: ScienceClaw is provided as a cloud service, so users do not need to configure local hardware environments such as GPUs. Only a browser with internet access is required. It is recommended to use modern browsers like Chrome or Edge for a full feature experience. Support for WeChat and Feishu plugins requires installation within the respective client applications.

2. Access the Official Website and Register: Open the ScienceClaw official website at https://scienceclaw.zidongtaichu.com/, and complete account registration and login. New users need to fill in basic information and verify their email or phone number. After the first login, it is recommended to review the platform's usage documentation and example projects to understand the functional layout.

3. Input Research Objectives: After logging in, submit a high-level research idea or project goal directly on the project creation interface. There is no need to break down steps or write detailed prompts. For example, you can input a vague goal such as "Develop a new perovskite material and evaluate its photovoltaic performance," and the system will automatically understand the research intent and initiate the planning process.

4. Confirm Project Planning: After the system automatically generates task breakdowns and research paths, users can view the task network structure on the planning confirmation page. This page supports switching between various preview modes, such as horizontal and vertical splits. Users can manually adjust task priorities, dependencies, and execution order. Once confirmed, click "Start Execution" to enter the autonomous execution phase.

5. Wait for Autonomous Execution and Intervene When Needed: The AI remains on-site for the long term, autonomously completing literature research, experiment execution, data analysis, and anomaly resolution, without requiring frequent user intervention. During the project's progress, users can随时 open the project dashboard to view intermediate results, evaluate research paths, hypotheses, and data, and make necessary judgments and corrections. The system will dynamically adjust subsequent tasks based on user feedback.

6. Obtain Research Outcomes: After the project is completed, the system automatically generates reusable research assets such as papers, datasets, models, code, and full experimental records. Users can view and download all outcomes on the project results page. All results are linked to the EviGraph evidence chain, and every conclusion can be traced back to its corresponding experiments and data.

7. Multi-Device Access and Advanced Usage: In addition to the web interface, the research tools can be accessed anytime and anywhere through WeChat and Feishu plugins. The official website provides detailed usage documentation, covering advanced features such as API interfaces, custom Agent configuration, and private data integration. Teams with customization needs are advised to refer to the documentation for in-depth configuration.

4. Pros and Cons Analysis

5. Comparative Analysis with Similar Tools

Selection Recommendations: For research teams that require full control over scientific projects and wish for AI to autonomously complete the entire process—from planning to experimentation to outcome consolidation—ScienceClaw is currently the most comprehensive choice in terms of functional coverage. It is particularly suitable for disciplines such as life sciences, materials science, and chemistry, which have high demands for iterative experimental processes. Its multimodal understanding capabilities and enterprise-level embodied execution support offer unique value in scenarios combining dry and wet lab experiments.

For early-stage research focused on hypothesis generation and experimental design, the hypothesis iteration mechanism of Google's AI Co-Scientist is worth noting. However, its inability to directly execute experiments means that researchers still need to perform a significant amount of verification work. OpenAI Deep Research is better suited for quickly obtaining comprehensive reports on specific research topics, while SciSpace focuses on literature reading and review scenarios. Neither involves experiment execution or data generation, making them suitable as supplementary tools for literature research rather than complete research platforms.

6. Editor's Summary

ScienceClaw demonstrates a clear technical differentiation in the field of scientific AI. Its core innovation lies in elevating the agent from a "task executor" to a "project manager," with the AutoProject engine achieving a full-cycle closed-loop process of planning, decomposition, execution, verification, repair, and knowledge accumulation. This directly addresses two long-standing pain points in scientific research: labor-intensive repetitive tasks and the difficulty of reusing research outcomes. From a technical architecture perspective, the hierarchical autonomous multi-agent system is more aligned with the actual collaboration methods of scientific research teams, while the EviGraph evidence graph framework responds to the critical concern regarding the credibility of AI-generated content at the mechanism level. The data showing a 40.74% improvement in traceable evidence rate compared to the optimal baseline indicates that this framework is not merely conceptual packaging, but a technically effective solution.

The platform's choice of a multi-modal foundation enables it to uniformly process heterogeneous scientific information such as papers, charts, experimental curves, and code. This capability holds practical value in disciplines like materials science and chemistry, which heavily rely on images and experimental data. Backed by the technical accumulation of the Chinese Academy of Sciences Institute of Automation and the continuous iteration of the ZiDongTaiChu large model, the platform has a solid foundation for long-term development.

This platform is suitable for the following groups: first, researchers who wish to be freed from repetitive experimental tasks; second, research teams that need to quickly validate multiple hypotheses; and third, scientific management institutions that pursue the standardization of the research process and the assetization of research outcomes. For individual researchers, the platform's project-level autonomous execution capability can significantly reduce the cost of managing multiple parallel tasks. For teams, the human-machine collaborative intervention mechanism ensures controlled research direction.

It is worth noting that as a new product, the platform's long-term performance in real, complex scientific projects still requires more case studies to validate. The depth of adaptation across different disciplines also needs more official data to be disclosed. Overall, ScienceClaw represents an important direction in the evolution of scientific AI from "auxiliary tools" to "research collaborators." Its project-level positioning and traceable evidence design provide a reference model for the industry.

7. Application Scenarios

• End-to-End Research Project Management: Conduct complete research projects from vague ideas to final paper delivery, including planning, experimentation, and paper writing. Researchers only need to input macro-level objectives, and the system will autonomously perform literature review, task decomposition, experimental execution, data analysis, and paper composition. This is suitable for exploratory research and rapid initiation of interdisciplinary projects.

• Literature Search and Review Compilation: Leveraging a library of hundreds of millions of interdisciplinary papers, the system can complete what used to take hours of literature search and review compilation in just a few minutes. It autonomously filters relevant literature, extracts key findings, and identifies research gaps, providing high-quality content for proposal writing, grant applications, and research background sections.

• AI Modeling and Experimental Iteration: Ideal for specific AI tasks such as steel surface defect detection and YOLO modeling, the system handles data processing, training, and iterative optimization. It automatically performs data preprocessing, model selection, hyperparameter tuning, and performance evaluation. In the event of experimental failure, it automatically traces the root cause and adjusts strategies, reducing manual trial-and-error costs.

• Life Sciences and Drug Discovery: Supports scenarios such as protein research, vaccine development, and drug screening, reducing drug property prediction cycles from one week to just 10 minutes. The system can simultaneously orchestrate specialized tools like molecular simulation, sequence analysis, and literature comparison, accelerating the candidate molecule screening and evaluation phases in early drug discovery.

• Materials and Chemical Research: Covers scenarios that integrate dry and wet experiments, such as compound simulation, materials metallurgy, and synthetic route planning. Multimodal capabilities enable the system to understand experimental images and curve data. The enterprise version can also connect with robotic arms and other devices for embodied execution, promoting the automation and closure of experimental workflows.

8. FAQ

Q: What is the fundamental difference between ScienceClaw and regular AI assistants (such as ChatGPT) in terms of research usage?
A: Regular AI assistants are centered around conversational interaction, excelling at knowledge answering and text generation, but are unable to independently advance a complete research project. ScienceClaw is a research-native intelligent agent platform, equipped with capabilities for project planning, task decomposition, long-term execution, evidence verification, and fault tolerance repair. It can independently manage the entire process from research concept to result delivery and accumulate reusable research assets.

Q: How does EviGraph evidence graph ensure the credibility of research conclusions?
A: EviGraph organizes the research process as a dynamic evidence chain: "Research question → Research gap → Hypothesis → Experiment → Discovery → Conclusion." Each conclusion is linked to corresponding experimental data and literature evidence. The system continuously verifies logical consistency across tasks, identifies deviations, traces root causes along the evidence chain, and autonomously repairs them. Combined with short-term and long-term graph libraries, it enables version rollback, ensuring conclusions are traceable and reproducible.

Q: Which academic disciplines does ScienceClaw support?
A: The platform currently covers eight major academic disciplines, including life sciences, materials science, chemistry, physics, and astronomy. It also provides cross-disciplinary literature support through its self-built library of over 100 million full-discipline papers. The compatibility of specialized toolchains may vary across disciplines; for detailed support capabilities in a specific field, please refer to the official documentation or contact the official team for confirmation.

Q: What hardware configuration is required to use ScienceClaw?
A: The platform is delivered as a cloud service, so users do not need to configure local GPUs or high-performance computing equipment. All features can be accessed using a browser with internet connectivity. For enterprise users requiring private deployment, hardware requirements depend on the project scale and concurrency level. Specific solutions should be obtained by contacting the official team for technical recommendations.

Q: How can researchers intervene in the AI's autonomous execution process?
A: The system is designed with dual modes: "Fully automated execution" and "Manual review." During project progress, users can open the project dashboard at any time to view intermediate results, evaluate and correct research paths, hypotheses, and data. The system dynamically adjusts the subsequent task network based on human feedback, ensuring the research direction remains under the control of the researchers.

Q: What content is included in the research outcomes generated after a project is completed?
A: After a project is completed, the system automatically generates a paper, dataset, model, code, and a full set of experimental records, forming a complete matrix of research assets. These assets are reusable and traceable, supporting direct invocation or iterative optimization in future research, thereby avoiding redundant work.

9. Project Links

• Product Official Website: https://scienceclaw.zidongtaichu.com/ (Provides product introduction, account registration, login access, and usage documentation)