A new study finds most AI agent platforms treat every task the same way, wasting CPU cycles in the process.
Researchers benchmarked three common agent tasks, retrieval-augmented question answering, web search, and software coding, measuring how latency interacts with CPU, memory, and disk I/O when multiple requests run at once. They found agent behavior varies wildly by task, and even the same tool produces very different resource patterns depending on what it's used for. Running requests concurrently exposed bottlenecks that differ by task, and counterintuitively, faster LLM responses or more CPU cores didn't always speed agents up. The team then built two optimizations, CPU-aware tool admission and task-aware CPU allocation, that match resources to what each task actually needs.
Most agent infrastructure today allocates resources uniformly, assuming a coding agent and a search agent behave the same way under load. This research shows that assumption has a real cost: task-aware allocation cut latency by about 5.4x for CPU-sensitive tasks and roughly 32% on average across the workloads tested.
It's a reminder that the agent boom is still running on infrastructure built for chatbots, not for programs juggling containers, disks, and APIs at once.