← Daily Brief for August 25, 2026
AgentX measures infrastructure using real agent-session behavior
Focus: Technical AI Engineering
Date: August 24, 2026
Topics: Agent evaluation, long context, inference infrastructure, KV-cache reuse, performance per watt
Evidence: Unspecified
Availability: Unspecified
Summary: NVIDIA published results using SemiAnalysis AgentX, an open-source InferenceX benchmark that replays recorded coding-agent sessions turn by turn. Unlike fixed prompt-and-response tests, AgentX preserves changing input and output lengths, accumulated context, reasoning time, tool-call latency, cache pressure, and varying concurrency. NVIDIA reports preview Vera Rubin NVL72 results of up to 30× more throughput per megawatt than GB300 NVL72 at 160 tokens per second per user.
Why it matters: Agent infrastructure cannot be evaluated realistically with a single fixed context length. Long-running agents create irregular bursts of model calls, tool waits, subagent work, and repeated context. A benchmark that preserves those trajectories is closer to measuring the actual cost and responsiveness of an agent harness.
Original commentary: This offers a valuable evaluation distinction for books and courses: measure not only output quality, but also completed-work latency, context growth, cache reuse, tool-wait time, concurrency, energy, and cost across the full loop.
Source: NVIDIA AgentX and Vera Rubin analysis