Gemini Flash via langchain-google-genai: thinking_budget=0 roughly halves tool-calling latency; parallel subagents cut a deepagents run ~5x
With langchain-google-genai 4.5 (init_chat_model(name, model_provider="google_genai", thinking_budget=0)), a single tool-calling turn on a current Gemini Flash model dropped from ~3.8 s (default, ~230 reasoning tokens) to ~1.8 s (no reasoning tokens) with the same, correct parallel tool calls; reasoning_effort="low" sat in between (~2.9 s). In a deepagents 0.7 orchestrator, wall time is dominated by sequential model turns, so the big wins were: (1) thinking off; (2) one batched tool that does N items per call instead of N calls; (3) a disk cache of deterministic tool results; (4) telling the orchestrator to issue independent task() calls in ONE message: LangGraph executes them concurrently, so two subagents cost one turn; (5) short subagent prompts with word caps and compact tool outputs. A 5-subagent run went from ~150 s / ~140k tokens to ~25 s / ~33k tokens. Measure tokens from LangSmith root runs (filter by your own run_id in metadata) rather than estimating.