Three agents, one spec, and a blind judging round that mostly proved they vote for themselves

Six agent sessions and one written spec. That is the entire apparatus: three builds from identical instructions, then three judging passes over the same anonymised trio ^[1][4][3]. No eval framework, no leaderboard.

The judging half is the part worth copying and the part worth distrusting. Two of the three judges put their own entry first without knowing it was theirs ^[10], a 2-in-3 self-preference rate across three votes ^[1], which leaves exactly one cross-vote in the whole run: Gemini's, for Claude ^[4]. As a ranking, that is unusable. As a control, it earns its keep. An agent asked to grade work it produced is, on this evidence, not a neutral reviewer, and that applies to the review step in your pipeline as much as to a contest.

The build artifacts are where the run actually pays. Codex wrote tooling to check that source files stayed under the maximum line count the specification asked for ^[8], which means a number in the spec became a mechanical test rather than a hope. Claude spent its session differently: 12 stages, procedural audio, a multi-phase DOH boss, accessibility and responsive work ^[7], reached through a loop of running tests, launching the game, screenshotting it and inspecting its own screenshots ^[12], and substantially more tokens than the other two ^[6]. Gemini delivered a working game with real Arkanoid functionality and much less of the verification and architecture ^[9][11].

The gap nobody measured is the one that shows up on an invoice. Claude's session ran past 20 minutes against Gemini's rough 5, at least a fourfold spread ^[2], and the author is straightforward that there was no stopwatch and these are observations rather than benchmark figures ^[5]. If you rerun this in-house, wall clock and token count per build are two columns that cost nothing to log and are the only ones the original could not report. They are also the numbers that decide whether the thorough agent is worth pointing at a small ticket.

The design choice that makes the whole thing legible is the refusal to help: no fixing mistakes, no "you forgot this feature", no second pass, so whatever the agent called finished was its submission ^[3]. That converts the exercise from a test of prompting stamina into a measurement of each agent's own threshold for done, which is what lands in a pull request on a normal Tuesday. The task selection follows the same logic, small enough for one session and broad enough to expose physics, architecture, UI, audio, controls and testing decisions ^[13].

With one task and one attempt each, nothing here settles which model is better. It settles that a comparison you can act on costs a written spec and an afternoon, and that the scoring half needs a judge with no entry in the contest.