Practical lessons on speeding up AI agents: inference speed (TTFT vs tok/s), tool call traces, building the tooling agents want, and hardware and sandboxing.