IBM, CoreWeave co-engineer controls for agent workloads

IBM Research's infrastructure needs changed as its model development practices evolved. Reinforcement learning added a task-execution stage: a checkpoint is loaded into inference, the model is asked to do something, and the result is measured. That stage changes what the underlying systems must accommodate, Belgodere said. Earlier, IBM Research built a large H100 cluster itself, acquiring the space and assembling it end to end — a task Belgodere described as huge.

The CoreWeave relationship grew out of that period. Much of the IBM Research cluster is single-tenant, with its own storage inside CoreWeave and additional capacity available within cost and security parameters. The joint engineering around identity management began with IBM supplying requirements for extending its internal identity systems into CoreWeave, then refining the implementation over several iterations. Belgodere framed the broader problem as a supply chain and provenance issue spanning hardware, firmware, kernel levels, code, data provenance, and agent images.

Where this lands may hinge on how well the security-performance tradeoffs Belgodere described hold up. IBM measures the performance impact of security controls against benchmark results, then discusses the tradeoffs with security teams. Belgodere also cautioned that early architecture decisions are expensive to undo — either through overbuying networking infrastructure or having to refit everything — so for teams standing up agent workloads now, the cost of getting the isolation and identity layers right the first time appears to be the practical question.