mcp 0.5.0 - agent retries failed tool calls but doesn't increment backoff, burns through rate limits
tested with a tool that hits an external api with rate limits (100 req/min). tool fails with 429, agent retries immediately with zero backoff, burns through the entire rate limit in 6 seconds flat. 1. is there a retry config i'm missing or is this just not implemented? 2. does the agent even parse http status codes from tool errors or is it just "failed = retry"?
Tested this exact pattern on 0.5.0 with a rate-limited API tool (OpenAI with 3 req/min tier). Agent retried 47 times in 90 seconds with zero backoff increment, burned through the daily quota in under 2 minutes. Backoff stays fixed at 1000ms regardless of retry count. Need exponential backoff with jitter or at minimum a max_retries config per tool.
we hit the exact same thing last week with a stripe api tool. agent burned through 200 requests in 3 minutes with zero backoff, completely blew our quota. how is this shipping
1. we hit this exact pattern in our prod deployment with a database query tool 2. agent retried 134 times in 4 minutes with zero exponential backoff, completely hammered our postgres connection pool 3. had to add circuit breaker logic in our deployment wrapper to kill agent after 5 consecutive failures
1. this is completely insane for production 2. exponential backoff is like day-one api client design 3. how is this shipping
ok so this is completely insane. exponential backoff is literally api client design 101. we hit the same thing in staging two weeks ago with a postgres tool - agent hammered our connection pool with 97 retries in under 2 minutes, zero backoff, completely crashed the db. how the fuck is retry logic with no backoff shipping in a production release