Building the enterprise environment for agentic AI
Intel Corporation
Anthropic
Intel shares five practical lessons for enterprises deploying agentic AI, emphasizing that success requires a complete environment beyond just LLM inference, with metrics like agent density per vCPU and P95 latency. The article outlines a three-phase approach: planning by agent density, adopting new observability practices, and scaling out by default.
Intel performed thousands of agentic AI workload experiments and derived five lessons for enterprise leaders. They argue that agentic AI is a larger systems problem, not just inference, and that most existing harnesses do not measure overall system performance. Capacity should be planned using agents per vCPU density, not agent count. Monitoring agent task latency (P95) is more informative than average CPU utilization. Systems hosting agents should default to scale-out, reserving scale-up for workloads with heavier per-agent compute. Intel extended Terminal-Bench to benchmark agentic AI workloads, using a deterministic record-replay of LLM responses to separate agent performance from LLM variability. The task mix included compilation, testing, database operations, and more. Deploying agentic AI should proceed in three phases: plan by agent density, adopt new observability focused on P95 latency, and scale out by default. The article also identifies enterprise personas for agentic AI: accountable leaders improving cycle time, productivity, and cost management.
- Сокращения
- vCPU = virtual Central Processing Unit — виртуальный центральный процессор
- LLM = Large Language Model — большая языковая модель
- P95 = 95th percentile latency — задержка 95-го процентиля
Source: MIT Technology Review —
original