cs.AIApr 14, 2026

SciFi: A Safe, Lightweight, User-Friendly, and Fully Autonomous Agentic AI Workflow for Scientific Applications

Authors: Qibin Liu, Julia Gonski

Organizations: SLAC National Accelerator Laboratory, 2575 Sand Hill Rd, Menlo Park CA 94025 USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Recent advances in agentic AI have enabled increasingly autonomous workflows, but existing systems still face substantial challenges in achieving reliable deployment in real-world scientific research. In this work, we present a safe, lightweight, and user-friendly agentic framework for the autonomous execution of well-defined scientific tasks. The framework combines an isolated execution environment, a three-layer agent loop, and a self-assessing do-until mechanism to ensure safe and reliable operation while effectively leveraging large language models of varying capability levels. By focusing on structured tasks with clearly defined context and stopping criteria, the framework supports end-to-end automation with minimal human intervention, enabling researchers to offload routine workloads and devote more effort to creative activities and open-ended scientific inquiry.

Figures & tables

Figure 1: Architecture of the SciFi agentic system, indicating the safety ensured by secure environmental operation, usability provided by a three-agent loop structure working under concrete "do-until" conditions with easy LLM swapping, and extensibility via a natural language interface.
Figure 4: Iteration counts for the agentic loop solving the basic experiment tasks ranging from file I/O, data processing and plotting to ML/GPU related, multi-step and complex environment setup tasks. Results are broken down into Simple, Structured, and Exhaustive task description levels.
Figure 5: Total walltime including execution time for the agentic loop solving the basic experiment tasks ranging from file I/O, data processing and plotting to ML/GPU related, multi-step and complex environment setup tasks. Results are broken down into Simple, Structured, and Exhaustive task description levels.
Task Desc.ItersWall (s)
Simple (NL)48532
Structured (ST)31334
Exhaustive (EX)1141422
Table 1: Summary of Experiment 3 Task 2 (completion of partial firmware implementation) performance for the Simple, Structured, and Exhaustive description levels.
Task Desc.ItersWall (s)
Rough (RH)922 †13 476 †
Simple (NL)20–66230–759
Exhaustive (EX)88898
† Sum of 7 sequential runs (TaskGroup memory); pass on run 7.
Table 2: Summary of Experiment 3 Task 3 (firmware and simulation interface design) performance for the Simple, Structured, and Exhaustive description levels.
Figure 15: Distribution of the reconstructed leading jet mass mJ1m_{\text{J}_{1}} of the R&D dataset (upper panel). The enhancement of the signal-over-background ratio (lower panel) with the cut of AD score reflects the performance of the AD algorithm.
Figure 16: Distribution of the reconstructed second-leading jet mass mJ2m_{\text{J}_{2}} of the R&D dataset (upper panel). The enhancement of the signal-over-background ratio (lower panel) with the cut of AD score reflects the performance of the AD algorithm.
Figure 17: Distribution of the reconstructed dijet mass mJJm_{\text{JJ}} of the R&D dataset (upper panel). The enhancement of the signal-over-background ratio (lower panel) with the cut of AD score reflects the performance of the AD algorithm.
Task Desc.ItersWall timeR&D AUC (signal 9.1%)BB1 AUC (signal 0.08%)
Open-ended114∼\sim 3.7h0.501 (fail)-
Interactive160∼\sim 9h0.8300.886
Guided46∼\sim 1.5h0.8540.885
Table 3: Performance comparison of different variants of agentic approach to the LHCO AD challenge.

Explore similar work

CardsList

  1. Experiments in Agentic AI for Science

    May 25, 2026Judy Fox, Geoffrey FoxDocument UnderstandingAI for Science

    School of Data Science, Department of Computer Science, and Biocomplexity Institute, University of Virginia, Charlottesville, VA, USA.

  2. Mimosa Framework: Toward Evolving Multi-Agent Systems for Scientific Research

    Mar 30, 2026Martin Legrand, Tao Jiang, Matthieu Feraud +5Autonomous Research AgentsMulti-Agent Systems

    Université Côte d’Azur, CNRS, ICN, Nice, France · Interdisciplinary Institute for Artificial Intelligence (3iA) Côte d’Azur, Sophia-Antipolis, France · Inria, Université Côte d’Azur, CNRS, I3S, France

  3. The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery

    Oct 13, 2026Robert Tjarko Lange, Luca Grillotti, Yutaro Yamada +2AI Agents for Scientific DiscoveryAI-Assisted Scientific Research

    Sakana AI