Four parallel agents greatly raised local AI throughput
The test ran through on one RTX 5090, with five requests assigned to each agent. Each request was limited to 1,024 tokens, and allowed up to eight tasks to run at once.
Combined rose from about 246 with one agent to 346 with two, 434 with four, and 534 with eight. Eight therefore delivered only about 2.2 times the combined of one agent, not eight times as much.
Because the 's computing power was divided among the agents, the average speed of each agent fell from about 257 to 67 . The gains became smaller at higher counts, so running several independent jobs together helped on this setup, but setting the agent count to the maximum was not automatically best.
Key points
- One agent produced about 246 combined .
- Four agents raised combined to about 434 .
- Eight agents reached about 534 , roughly 2.2 times the one-agent result.
- Per-agent speed fell from about 257 to 67 as more agents shared the .
- Adding agents beyond five produced increasingly small gains.