Introducing AgentX: InferenceXs new agentic inference performance benchmark. (1/2)🧵

(2/2)

another benchmark, because apparently agents need report cards before doing the job

Goated first few seconds, had to watch 3 times.

we love it

에이전트엑스라는 에이전트 추론 벤치마크를 새로 냈다고 합니다.

Bigger than ranking GPUs: AgentX has driven 50+ PRs across vLLM, SGLang, Dynamo and the KV stack, targeting routing, offload and scheduling. A shared workload can turn each fix into lower session latency and cost—making coding and enterprise agents cheaper to serve at scale.

Every new agentic benchmark measures inference speed and, rather more precisely, the priorities of whoever chose what to put on the clock.

when you replay claude code sessions, are you benchmarking the model or the model plus harness?

Closed-loop adds a compositional confound. For any per-request metric m, m̄_A−m̄_B = Σ_w p_A(w)[m_A(w)−m_B(w)] + Σ_w[p_A(w)−p_B(w)]m_B(w). The second term is mix drift because A progresses farther in 1h. Could AgentX report both raw and common-p* standardized gaps?

Agentic inference is the first benchmark that will actually price power. Tokens/sec without joules/token is a vanity metric. Publish both.

benchmark is just a fancy word for 'how fast can we make our model obsolete