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

The Headline Bug: A Whole Column of Data Didn't Exist

CPython 3.11+ limits int→str conversion to 4,300 digits by default (a DoS mitigation). My Python agent stringified each prime it found:

mersenne_primes.append(str((1 << p) - 1))

The 24th Mersenne exponent is p=19937, and 2^19937−1 has 6,002 digits. So for any request of 24+ primes, the tool raised ValueError - and the A2A response dutifully delivered the stack-trace text instead of a result:

"text": "Exceeds the limit (4300 digits) for integer string conversion; use sys.set_int_max_str_digits() to increase the limit"

The benchmark's Python column was structurally incapable of producing data at N≥24. The kicker: the stringified list was never used. The tool returns only elapsed_time. The fix is deleting the str() - which also removes formatting work from the timed region that the Node and Rust agents never paid (Go had the same dead val.String() call).

Fix: PR #1 - plus a switch from time.time() (wall clock, non-monotonic) to time.perf_counter(), and a regression test at count=24.

Before/after, N=24 row:

The Harness Was Reading Its Data from LLM Prose

The Python agent's timing came back in two places: a structured elapsed_time in the tool artifact, and the model's prose. The harness regexed the prose first:

m = re.search(r"It took ([\d\.\-e]+) seconds", text)

In live runs, Gemini said "Calculating the first 5 Mersenne primes took 4.9591064453125e-05 seconds." one time and "The calculation took 2.40715261301375 seconds." the next. Neither matches "It took". Every datapoint survived only because a fallback happened to dig out the structured value. Measurement data should never depend on how an LLM felt like phrasing a sentence.

Meanwhile Node and Rust reported elapsed.toFixed(2) ms - so sub-10µs runs returned 0.00ms, which parses to 0.0 and silently vanishes from a log-scale chart.

Fix: PR #2 - structured artifact first, prose as last resort; 4-decimal timing output.

The Agents Lied About What They Computed

The exponent table has 26 entries, but both direct agents echoed the requested count:

$ curl ... "Calculate the first 100 Mersenne primes"
node: "Found first 100 Mersenne primes in 4450.78ms."  # computed 26
rust: "Found first 100 Mersenne primes in 2160.45ms."  # computed 26

Fix: PR #3 - report primes.length / primes.len().

The Chart Compared Two Different Universes

Python and Go requests route through Gemini tool-calling; Node and Rust regex the number out of the message and compute directly. The RTT chart presented all four as one comparison "including LLM/Tool calling" - true for exactly half the lines. Median RTT: Rust 2.6ms, Node 4.6ms vs Go ~1.6s, Python ~1.8s. That ~400× gap is pipeline architecture, not language performance.

Fix: PR #4 - agents tagged by pipeline; the chart renders direct vs via Gemini LLM as distinct series on a log scale, with a title that says what's actually compared.

Validation Found Two More (My Favorites)

The fix made the code too fast for its own benchmark. With formatting removed from Go's timed region, small-N runs dropped under a microsecond - and Go's time.Duration started printing 836ns, a suffix the harness parser had never seen. The datapoint silently became N/A. One more parser branch fixed it.

Gemini refused to repeat itself. My harness reused deterministic contextIds, and ADK keeps per-context session history. On a rerun, the model answered "I already did that. Do you want to do it again?" - without calling the tool. Three datapoints gone. Benchmark IDs are now unique per run.

Result

96/96 datapoints, up from a baseline where one language column crashed, small-N values were unplottable zeros, and two datapoints per rerun depended on an LLM's patience. Nine bugs, four PRs, one afternoon of reproduction scripts - all archived with before/after charts.

How Google AI Fit In

The suite runs on ADK + gemini-2.5-flash tool-calling end to end. Two hard-won lessons for anyone building Gemini agents: read structured tool artifacts, never model prose, when a machine consumes the output - and never reuse context IDs for independent requests unless you want session memory changing your results.

Disclosure: I used Claude Code as the debugging/automation agent for this work - it reproduced each bug, wrote the fixes and regression tests, and ran the before/after benchmark sweeps. Every bug, number, and PR above is real and verifiable in the repo.