My benchmark's Python column was N/A for a year — CPython's 4300-digit limit, and eight other bugs

Short summary

A structured breakdown of nine bugs in a multi-language A2A benchmark suite, each with a fix PR. The primary bug was CPython 3.11's 4300-digit int→str limit crashing the Python agent on large Mersenne primes, but the stringified result was never consumed. Additional bugs included fragile LLM-prose regex parsing, sub-microsecond timing rounding to zero, misleading RTT chart labels, and Gemini refusing to repeat work due to reused context IDs.

• •CPython 3.11's 4300-digit limit crashed Python agent at N≥24; fix was deleting unused str() call
• •Timing data depended on regex-matching Gemini's prose phrasing, which never matched the expected pattern
• •RTT chart conflated pipeline architecture (direct vs Gemini-routed) with language performance, creating a misleading 400× gap

Generated with AI, which can make mistakes.