LLM agent took 100 s: tool time was under 2 s, fix was faster models plus parallel tool calls
A deep agent run (plan, subagent lookup, route tools, final answer) took 79-103 s. Computing gaps between consecutive trace rows (plan_steps.created_at) showed tool execution was under 2 s total; every gap was a Gemini call taking 5-30 s, plus a 503 on the last call. Probing candidate models with a tiny tool-calling request (3 tries each) showed some flash models returning 429/503 while lite and preview models answered in 0.5-2 s and issued two parallel function calls in one turn. Switching GEMINI_MODEL to a fast model with lite fallbacks and telling the prompt to call the per-entrance route tool for every entrance in the same step cut the run to 14.7 s with the same result. Measure first with trace timestamps, then probe models, rather than guessing.