The Next Wave of Confluence: Agentic Workflows in HPC and Cloud

The conversation around HPC and Cloud has centered on bursting for capacity, cloud suitability for HPC workloads, data management, and cloud-like capabilities for on-premise systems. Today, the driver has fundamentally shifted: Artificial Intelligence is reshaping not only the hardware landscape but the scientific method itself. The rise of "AI for Science" is creating workflows of unprecedented complexity and capability. We see this in the emergence of:

- Scientific Foundation Models: Large models trained on vast corpuses of scientific data (e.g., proteins, climate data, material properties) that can be fine-tuned for specific research questions.

- The Agentic Developer: Tools like Gemini CLI, Claude Code, and multi-agent orchestrators like Gastown are changing how HPC software is maintained. Agents are now capable of refactoring legacy code and generating boilerplate, but they require robust "guardrails".

- Agentic Co-Scientists and their federation: AI agents such as Google’s Co-Scientist demonstrate how AI can drive hypothesis generation, develop execution plans and even use tools. Middleware like Academy (from Globus Labs) is demonstrating how "Federated Agents" can span on-premise supercomputers and public clouds, handling asynchronous execution and decision-making while respecting site-specific boundaries.

- Standardized Interfaces: The emergence of the Model Context Protocol (MCP) offers a path to standardize how AI agents interact with HPC schedulers, filesystems, and data services, moving us away from brittle, ad-hoc scripts.

This shift presents critical questions for our community, which we will explore through the lens of key personas:

- Supercomputing Centers: How do we support "Agentic" users without compromising security? If an agent running in the cloud needs to steer a simulation on-prem, how do we implement mechanisms that comply with center policies?

- Cloud Providers: How can we facilitate hybrid infrastructure that bridges the rapid AI progress of cloud vendors with the specialized, secure nature of on-premise scientific infrastructure?

- Scientific and Industrial Users: How do we keep the human in the loop? As we adopt tools like AlphaEvolve or autonomous coding agents, how do we design workflows that enforce human checkpoints for verification before large-scale resources are consumed?

- Hardware and Software Vendors: How do we expose infrastructure (schedulers, data movers) via MCP servers so that the next generation of AI tools can discover and utilize them securely?

This BoF will move beyond the "if" to the "how," focusing on the practical challenges and emerging solutions in this new, AI-driven era of HPC.

---

Organizers:

• Felix Schuermann