Check whether LLM 'slow call' timeouts are per-request or provider-wide bursts before raising timeouts
Multi-agent deepagents run on Gemini (langchain-google-genai, 12 s client timeout which the API enforces as a ~10 s server deadline, one immediate retry). Live end-to-end tests failed ~50% of runs with a subagent timing out twice in a row. My first hypothesis was that the long drafting agent needed more time, and I recommended a 25 s timeout for it. Pulling every LLM run of the last 50 minutes from LangSmith (client.list_runs by trace, filter r.error, group by lc_agent_name and start time) disproved it: 61 completed drafting calls had max 8.4 s, and the 504 DEADLINE_EXCEEDED errors clustered within a few minutes across all agents, including an orchestrator whose calls normally take 2-3 s. So these were provider-side stall bursts, and the immediate retry landed in the same burst. Better mitigations: retry on a different provider (failover) rather than the same one; let a subagent that gives up end only its own task (code fallbacks cover its output) instead of raising through the orchestrator and failing the whole run; and record per-run stalls so test failures can be attributed. Also: a LangSmith 'last N minutes' scan over many traces is slow (minutes); run it in the background.