Speaker
Description
We present open-source agentic AI frameworks, RooAgent and FCCAgents, that apply large language models (LLMs) to automate HEP analysis workflows through natural-language interaction, while keeping all physics computation within established, validated software.
RooAgent exposes ROOT-based analysis operations — histogram inspection, event selection, kinematic-distribution visualisation, fitting, and significance estimation — as discrete tools that an LLM agent invokes in response to plain-language prompts. The analysis logic is implemented in PyROOT, and the agent is responsible only for selecting tools and supplying their arguments, so that the results remain traceable. Two interchangeable backends are supported: a LangGraph-based agent and a Model Context Protocol (MCP) server. We validate the system on Monte Carlo samples of pp → ZH (Z → ℓ⁺ℓ⁻, H → b̅b), multi-task signal–background workflows, a toy statistical analysis, and ATLAS open data for H → ZZ* → 4ℓ.
FCCeeAnalysisAgents are implemented using Claude Code. It extends the approach to FCC-ee analyses within the FCCAnalyses framework. It provides sub-agents spanning process selection, FCCAnalyses script generation, BDT training, analysis review, and analysis note drafting. A planner agent selects signal and background processes, encodes constraints, and provides a machine-readable JSON plan that drives all downstream stages. We demonstrate the full workflow for a physics analysis at FCCee.
| I am the presenting author | Yes |
|---|