Intel says enterprise agents should be sized by vCPU density, not agent count
MIT Tech Review AI · rss · 2026-07-27
Intel argues that enterprise agentic AI is a systems problem, not just an inference problem: agents need CPU capacity, resilient data access, policy-aware tool use, observability, memory, and predictable scaling.
Five practical lessons from Intel’s experiments
- Size fleets by agents per vCPU, not raw agent count.
- Track task latency instead of only average CPU utilization.
- Use scale-out by default; reserve scale-up for heavier per-agent compute or architectural constraints.
- Measure the full workflow, not just the model.
- Build around workflows that already have rules and service levels, such as code generation, regression testing, ticket triage, market analysis, and security review.
Benchmarking approach
Intel extended Terminal-Bench with profiling, telemetry, and replay so agent performance could be separated from LLM variability.
Why it matters
The article frames production agent systems as an orchestration and infrastructure challenge: the real goal is reliable business workflows that scale under governance and cost constraints.
More from coding & agent
- MediaCrawler tops 57,937 stars as a Python crawler for major Chinese content platforms — NanmiCoder · 2026-07-27
- Dear ImGui reaches 75,115 stars as the lightweight C++ UI library keeps growing — ocornut · 2026-07-27
- Shenzhen’s new humanoid robot fighting league scores taunts and knockdowns — aitrendz_xyz · 2026-07-27
- A DeepSeek-powered oil-trading agent now bets daily on Polymarket — 0xJeff · 2026-07-27
- Agent workspace visualizer will ship as a plugin with custom layouts — jasonkneen · 2026-07-27
- Agentic AI only works at scale with orchestration, latency control, and cost discipline — nordicinst · 2026-07-27