IBM and CoreWeave co-design controls for agent workloads

IBM and CoreWeave are collaborating to address the challenge of agent workload isolation in increasingly complex computing environments. As reinforcement learning systems move beyond model training, they require more robust infrastructure to support various workloads interacting with tools, storage, and services. Brian Belgodere, a senior technical staff member at IBM Research, explained that the reinforcement learning process involves taking a model's checkpoint, loading it into inference, and measuring its performance in a testing phase.

This shift in workload demands led IBM to partner with CoreWeave, which built a large H100 cluster to support IBM's research needs. The collaboration has expanded to include joint engineering efforts on identity management and workload controls, with IBM providing requirements to extend its internal identity systems into CoreWeave.

IBM's cluster is largely single-tenant, with its own storage deployed within CoreWeave, and additional capacity available within cost and security constraints. The partnership also includes CoreWeave Sandboxes, which offer isolated execution environments for researchers to choose their preferred resources and access levels. Belgodere emphasized the importance of measuring the performance impact of security controls and the need to consider the broader supply chain, including hardware, firmware, kernel levels, code, data provenance, and agent images.