I participated in a panel discussion at the AGENT4SC’26 workshop on Agentic AI for Large-scale Science at IEEE eScience 2026 in Naples, Italy, on September 28, 2026 where we grappled with some of the most challenging questions facing agentic systems in scientific research.

The discussion centered on several provocative topics: If agents design experiments, run workflows, and draft papers, what does scientific authorship mean when humans primarily verify rather than create? We debated whether the scientific community can keep pace with AI-generated claims—production is cheap, but verification is not. The panel questioned whether having agents review other agents’ work constitutes real verification or simply automation of the appearance of oversight.

We also explored deeper infrastructure concerns: whether perfect provenance traces explain decisions or just provide detailed alibis, and whether building national research infrastructure on proprietary AI models creates dependencies we’ll regret. A recurring theme was the gap between generation and verification—agents produce code, insights, and scientific claims faster than humans can validate them. The talks also focused on detecting silent failures in multi-agent systems, provenance tracking, and evaluating scientific tool interfaces.

Workshop website: https://agent4sc.github.io