Blog

Your Agent Can Provision GPUs. Should It?

A training task led an agent to request an H100 node pool. Who should approve the resource commitment?

Armor12 min read

A training job needs compute. A coding agent has cloud credentials. Somewhere between those two facts sits a budget decision.

In one customer development environment, an agent checked accelerator quotas, issued a command to create an H100-backed node pool and prepared a Kubernetes training job. The reviewed configuration began at zero nodes and capped the pool at one.

The finding raises a practical question: how does an organization decide which infrastructure commitments an agent can make while completing a task?

The quiet cost of “make it work”

A cluster configuration may be just another step to an agent. To the platform team, it can create capacity, consume quota and incur charges. A zero-node pool may still lead to a node being scheduled later. Whether it does so depends on autoscaling, quotas, scheduling and the provider's response.

The right investigation follows the whole chain: request, cloud operation result, node state, workload scheduling and deletion. A create command alone answers only the first part.

Put spending authority next to technical authority

Give agents a bounded project or account, with machine types, regions, maximum capacity and permitted lifetimes set by policy. For expensive resources, require a budget owner to approve a request that specifies the workload, maximum size and duration.

Approval should be enforced outside the agent's instructions. Quotas and policy checks can prevent an unapproved request from becoming infrastructure; budget alerts are useful but may arrive after costs start. Apply expiry and cleanup rules that an independent controller can enforce.

Questions for the owner

  • Did a node pool or an actual node get created? For how long?
  • Was the machine type and maximum capacity within the task's authorization?
  • Did the training job run, and did cleanup leave any resources behind?
  • Can the same identity expand the cap or remove its own expiry?

A GPU request is not inherently suspicious. It becomes a governance problem when nobody can tell which resource decision the agent was allowed to make.

agentscloud-securitygovernance

See what your own fleet is running

Armor1 discovers and risk-scores every MCP server, skill and tool your developers have connected. Free for personal use, no credit card.

Start Free