Agents load testing results

[jennsun]

Across all 5 runs (20 config-runs each):
┌─────────┬──────────────┬─────────┐
│ Compute │ Avg Peak QPS │ Avg QPS │
├─────────┼──────────────┼─────────┤
│ Medium │ 123.5 │ 45.5 │
├─────────┼──────────────┼─────────┤
│ Large │ 278.0 │ 100.1 │
└─────────┴──────────────┴─────────┘

Large is ~2.2x medium on both peak and average QPS.

Recommendations:

• Medium: 2 workers — clearly the best, 33% higher peak QPS
  than 4 workers. More workers on medium compute likely
  causes contention on the limited CPU/memory, so the
  overhead of managing extra processes outweighs the
  parallelism benefit.
• Large: 6-10 workers — the differences are small (268-288,
  within ~7%), so any of these work. 10w edges out slightly
  but it's within noise. I'd recommend 8 workers as a safe
  default since it's in the middle of the plateau and avoids
  the slight dip at 12w.

The takeaway: on smaller compute, fewer workers is better
(less overhead). On larger compute, there's a sweet spot in
the middle — adding workers beyond that introduces
diminishing returns and eventually contention.

So it looks like on average we’re seeing large compute with ~2x medium on peak (278 peak avg vs 123.5 avg) and average QPS, but more workers is not necessarily always better