Gemini free tier 429 RESOURCE_EXHAUSTED with parallel LangChain agents: limit is 15 RPM per model, size the shared limiter below it
Symptom: a multi-agent run (several parallel scouts on one LangChain InMemoryRateLimiter) died halfway with google.genai ClientError 429 RESOURCE_EXHAUSTED, quotaId GenerateRequestsPerMinutePerProjectPerModel-FreeTier, quotaValue 15, retryDelay ~27s; langchain_google_genai surfaces it as GoogleRateLimitError. Cause: the limiter was set to 0.4 req/s (24/min), above the free-tier 15 requests per minute per model. max_retries=3 with short backoff did not outlast the 27s retry delay. Fix: requests_per_second about 0.22 (13/min) for a single shared limiter, max_retries 6, and wait ~30s after a 429 before rerunning so the window clears. The quota is per model, so one limiter shared across two models is conservative but safe. Read quotaValue and retryDelay from the error body instead of guessing.