AutoResearch

AutoResearch, developed by EvoMap’s Infinite Evolution Lab, is an open-source agent workflow tailored for AI and machine learning research. Instead of treating research as a simple text-generation task, it structures the process into a multi-stage execution pipeline that converts ideas into paper-ready evidence packages. ### Architectural Breakdown The framework operates through a multi-step workflow designed to cover the standard research lifecycle: * Direction Discovery & Planning: The system accepts user-defined research ideas or automatically discovers promising research directions by monitoring recent papers, developer communities, and open-source trends. Once a direction is locked in, it generates experiment plans and writes implementation code. * Review & Execution: The generated codebase undergoes a review phase prior to execution. During active runs, the framework captures execution metrics, system logs, and explicit failure causes if an execution errors out. * Analysis & Independent Evaluation: Once execution finishes, the pipeline runs result analysis and routes the outcomes through an independent evaluation layer that generates critic reports and blind reviews. ``` Idea Discovery / Input ──> Planning & Code ──> Review ──> Execution ──> Analysis ──> Blind Review ``` ### Statefulness and Disk Persistence A frequent issue with multi-agent orchestration frameworks is losing state during long execution loops or pilot failures. AutoResearch solves this by maintaining a stateful, recoverable architecture where all pipeline outputs are written directly to disk. The system continuously persists: * Experiment plans and source code files * Execution logs, quantitative metrics, and explicit failure diagnostics * Critic reports and blind review evaluations This local storage model guarantees complete visibility and human-in-the-loop control. Developers can inspect generated artifacts on disk at any stage, stop execution, manually modify code or configurations, and resume the run. Furthermore, the pipeline is designed to iterate recursively—using pilot results, critic feedback, and execution rounds tracked in its project monitor to adjust subsequent runs. ### Environment & Specifications * Language Runtime: Python 3.10+ * License: Apache-2.0 * Research Reference: arXiv:2608.17906 ### Developer Perspective AutoResearch provides a transparent approach to automated ML research workflows. By emphasizing disk-persisted state, human takeover points, and independent evaluation loops over black-box execution, it serves as a practical, controllable utility for running structured ML experimentation pipelines.