Preallocate agent sandboxes speculatively: P99 latency down 2.9x and peak memory down 45.9%
arXiv 2607.23933·low signal
SpecBox schedules sandbox creation speculatively rather than on demand when serving LLM agents, reporting up to 2.9x lower P99 latency and 45.9% lower peak memory. This is an infrastructure-side lever for anyone whose agent loop pays container-startup cost per tool call. Single-source and I read the abstract only, so treat the magnitudes as claimed rather than confirmed.