Exploring the sweet spot of AI agent parallelism and why more concurrent tasks often lead to slower execution.

How a subtle race condition in file-based rate limiting allowed unlimited API calls, and the atomic operations that fixed it.