Multi-Agent Parallelism Throughput Benchmark Tested
BringTea_666 · reddit · 2026-07-12
This post shares a real-world benchmark for OpenCode / agent parallelism: running Qwen 3.6 35B via LM Studio on an RTX 5090, scaling the number of parallel tasks from 1 to 8 to observe changes in throughput, efficiency, and VRAM pressure.
Main Conclusions
- Total throughput does not scale linearly: 8 agents yield a combined throughput of only about 2.2x that of a single agent, not 8x.
- Significant per-agent speed drop: Individual agent speed drops from roughly 256 t/s down to 67 t/s.
- Diminishing marginal returns: Beyond 5 agents, adding more concurrency yields minimal benefits while resource consumption continues to climb.
Author's Recommendations
- The sweet spot is 4–5 agents:
- 4 agents: approx. 434 t/s, offering well-balanced efficiency.
- 5 agents: approx. 470 t/s, approaching peak total throughput.
- Beyond 5 agents, VRAM pressure and context/KV cache overhead escalate rapidly, with negligible throughput gains.
Additional Observations
- Each agent consumes its own full context window.
- During concurrency, KV cache and compute partitioning are the actual bottlenecks.
More from coding & agent
- HeyGen adds a media-sourcing skill for coding agents with 75k images and 10k tracks — HeyGen · 2026-07-22
- Agent search bottlenecks are now about variance, not raw latency — rohanpaul_ai · 2026-07-22
- LangSmith adds tracing for Pipecat, LiveKit, OpenAI Realtime, and Gemini Live — LangChain · 2026-07-22
- An MCP server signs every AI agent tool call into a verifiable Merkle chain — Funky_Chicken_22 · 2026-07-22
- Annotated transcript of a Claude Code team interview is now available — trq212 · 2026-07-22
- Claude Code skill uses 10 Markdown rules to make outputs ADHD-friendly — alex_verem · 2026-07-22