Open Code parallel agents benchmark on RTX 5090 with Qwen3.6 35B

A user benchmarks Open Code's multi-agent throughput using an RTX 5090 and Qwen3.6 35B via LM Studio, finding that running at least four parallel agents prevents losing half of the potential performance.

• Single agent achieves 245.77 combined tokens per second (t/s), while eight agents yield only 533.90 t/s, representing a 2.2x increase rather than linear scaling.
• Individual agent speed drops sharply from 256.54 t/s to 67.22 t/s as GPU compute is divided among tasks.
• Peak efficiency of 70.4% occurs at two agents, but drops to 27.2% by eight agents due to KV cache and overhead costs.
• The bottleneck is identified as KV cache expansion (8x for eight agents) rather than context size splitting.

The author recommends setting parallel agents to four or five in Open Code to maximize throughput gains without excessive VRAM overhead or diminishing returns.