Making agents fast – what I've learnt over the last 12 months

Martin Alderson, drawing on a year of optimising agentic workflows, argues that inference speed is only part of agent latency. By his own rough illustration, halving inference time cuts about 20% off a task when tools, provisioning and suspend/resume stay unchanged. He ranks thinking budgets as the first thing to check, then tool-call traces, slow shell startup, polling loops, disk and CPU choice, and microVM resume. The figures are illustrative and unsourced, but the checklist is practical for anyone running agents in sandboxes.