ASRP: An Open-Source Research Framework for AI Agent Collaboration Following Scientific Methodology

ASRP (AI-agent collaborative Scientific Research Protocol) is an open-source framework that encodes scientific methodology into AI agent workflows, aiming to resolve the conflict between speed and rigor in AI-assisted scientific research. It supports at its core mechanisms like pre-registration of experiments, independent cross-validation, audit trails, and reproducible research processes. It has produced 20 theoretical physics papers in 16 days, and also provides a desktop application to lower the usage threshold. It is built on the OpenClaw general framework as a dedicated skill layer for the scientific field.

ASRP originated from real scientific research practice in March 2026: a researcher collaborated with two AI agents, producing 20 theoretical physics and mathematics papers in 16 days. 14 of them were submitted to journals like Physical Review D, and 7 are under review. During the process, the agents actively identified and corrected methodological errors, but also exposed the risk that high-speed output might be accompanied by systematic biases. The ASRP framework was thus refined, encoding the core principles of scientific methodology into AI agent workflows to ensure efficient and rigorous collaboration.

ASRP is built around five key mechanisms:

1. Pre-registration of Experiments: Mandates registration of hypotheses, methods, and analysis plans before research commences to prevent post-hoc hypotheses and selective reporting;
2. Independent Cross-Validation: Different agents perform separate validation to ensure result reproducibility;
3. Audit Trails: Records all decisions, data accesses, and corrections to comply with academic norms;
4. Token Budget Management: Allocates suitable models to different roles (e.g., Claude Opus for theorists, Gemini Flash for literature retrieval) to optimize costs;
5. Separation of Discovery and Validation: Isolates the permissions of theorists and reviewers to guarantee the objectivity of validation.

ASRP provides a desktop application built with Electron + TypeScript, featuring a four-step setup wizard, real-time dashboard, assistant chat panel, agent management, file manager, paper pipeline, experiment registry, and Ollama integration (with local Gemma support). Technically, ASRP is a domain-specific skill layer for science on top of the OpenClaw general framework: OpenClaw handles tool orchestration, while ASRP focuses on enforcing scientific methodology, allowing researchers to concentrate on scientific problems.

The project's examples/portfolio directory records the complete analysis of the founding case. The 20 papers cover: fine-structure constant α-related (4 papers), Riemann hypothesis-related (2 papers), superconductivity (3 papers), membrane models (2 papers), quark-lepton-prime mapping (1 paper), and mathematical physics (8 papers), demonstrating the framework's versatility and adaptability.

ASRP is currently in the Alpha phase (v0.1.0). The desktop application uses demo data, and the core framework is operational but agent integration needs improvement. Future directions include: supporting more discipline-specific templates, deep integration with academic databases and preprint platforms, multi-person collaboration support, and more powerful visual analysis tools.

The significance of ASRP lies not only in improving research efficiency but also in establishing followable norms for AI-assisted scientific research. In today's era of rapid AI capability iteration, we need mechanisms to ensure research quality. ASRP reminds us: the essence of science is the pursuit of truth, and this core mission remains unchanged regardless of tool evolution. When AI agents become regular members of research teams, ASRP will be the necessary infrastructure to maintain scientific rigor.