Part of the TPC Seminar Series

---

Abstract:

Agents are often designed around prompts, tools, skills, and delegation. These describe how an agent approaches work, but not the work itself. Reliable scientific progress requires a durable, structured environment in which domain objects, executable operations, outputs, and evidence have explicit identities and relationships.

I will explore this distinction through examples from coding agents, biological analysis, and network-based biomodelling. In coding, the relevant objects include source files, plans, tests, probes, and build artifacts; in biology, result tables and analysis tasks; in biomodelling, knowledge bases, networks, perturbations, and simulation runs. Across these settings, the same questions recur: What is the unit of work? How is it persisted and recomputed? How can an agent invoke operations, and how can the results be inspected and verified?

I will present a design built around persistent workflows, composable tasks, inspectable artifacts, and provenance, with an evidence layer that distinguishes observations and computed results from claims and hypotheses. This separates the durable system from replaceable agent strategies such as skills, planning, and delegation. The goal is not to force every domain into a universal ontology, but to build scientific environments in which agents can make reproducible, evidence-grounded progress.

Biography:

Miguel Vázquez leads the Genome Informatics group at the Barcelona Supercomputing Center (BSC). His work focuses on building computational infrastructure for large-scale bioinformatics, combining data integration, biological knowledge bases, and reproducible workflows to support biomedical research. He is the creator of Rbbt/Scout, an open-source bioinformatics toolkit, and is developing the Scout ecosystem to bring these foundations into AI-driven scientific work. His current work explores how AI agents can operate on persistent computational workflows, use domain knowledge, and produce inspectable results with traceable provenance, bridging scientific software engineering, computational biology, and agentic AI.