Globus
Data movement, remote execution, and cross-site automation
Globus is supporting infrastructure, not a peer alternative to URSA, MADA, or Academy.
Use:
Globus Transfer to move large datasets reliably between authorized endpoints.
Globus Compute to execute approved functions on remote resources.
Globus Flows to coordinate transfer, compute, and approval steps.
Choose Globus when data and compute span institutions or facilities. Pair it with an agent framework only when an agent must decide what to transfer, execute, or monitor. Keep authorization, endpoint ownership, data-use agreements, and retry/idempotency behavior explicit.
Additional Resources
ExaServe
Self-hosted inference on HPC allocations
ExaServe is a model-serving layer, not an agent framework.
Use ExaServe when the team must serve a model from an HPC allocation and has:
A supported computing site and scheduler configuration.
An allocation and approved model weights.
Staff who can operate Ray, the inference engine, networking, proxies, and site-specific patches.
A plan for authentication, authorization, rate limits, monitoring, model governance, and teardown.
Use MAG instead when an approved centrally managed model endpoint meets the need. A plain HAProxy or unproxied ExaServe deployment does not add authentication by itself; use an approved authenticated front end and site security controls before exposing a service.