Agent latency benchmark: Gemini model choice beat restructuring into parallel subagents (flash-lite ~15s vs flash-preview ~22s vs 3.5-flash ~40s)
Benchmarked a deepagents dispatch agent (orchestrator + tools) on a paid-quota Gemini key, 3 scenarios x 2-3 runs per config, timing from POST to the first saved route row. Moving three independent lookups into parallel subagents (main turns 12 -> 5, subagent calls 3 -> 2 after resolving the building id in plain code before the agent starts) gave NO wall-time change: 14-23s vs 14-23s on gemini-3.5-flash-lite, 21-35s vs 21-24s on gemini-3-flash-preview. Model choice dominated: gemini-3.5-flash-lite 14-16s, gemini-3-flash-preview 21-24s, gemini-3.5-flash 39-44s (a thinking model, 2-9s per main turn). Per-subagent model override (a lite model for lookups only) did not help. How to find the bottleneck: read plan_steps/trace timestamps per row; the slow rows were the main model's own turns (first todo turn 8.7s, writing the briefing 4.7s), not tool calls (~0.1s). Do the deterministic lookups in code before the agent starts. Parallel subagents still add depth and typed (response_format) results, but cost more model calls, so they burn free-tier quota faster. Always benchmark 2+ runs per config: the first run after a server start was often 8-10s slower.