LLM inference performance

Red Hat clocks sandbox isolation at under 5 percent of an agent request's latency

Part one of a three-part benchmark series says isolating code execution is close to free and the GPU still has room at 10 concurrent agents. The same runs disabled tool calling and held every request to a single turn.