Speaker
Description
Recent coding agents can execute increasingly complex scientific workflows, but realistic high-energy physics (HEP) analyses remain nonlinear and require domain expertise, systematic validation, reproducible evidence, and human scientific judgment. We present a human-governed, skill-oriented workflow for agent-assisted HEP analysis. Native coding agents perform planning, coding, computation, validation, and documentation, while reusable Skills provide HEP-specific methods and checks on demand. Rather than imposing a dedicated orchestrator or fixed workflow graph, the approach builds on native agent capabilities and composes Skills according to the task and its risks. Human researchers retain scientific authority and review versioned evidence and analysis packages through GitLab. We demonstrate the workflow through a joint measurement of Rb, Rc, and Rs using LEP Open Data, and introduce evaluations ranging from Skill activation tests to blinded full-analysis benchmarks derived from real analysis history. These studies explore how agent autonomy, domain knowledge, and human oversight can be combined for reliable and reproducible HEP analysis.
