Die Gauntlet-Schleife: Iterativer Agent-Workflow mit blinden Kritikern

Step 1: Set a real, inspectable bar (e.g. 'most recent Call of Duty' level, not 'a great FPS'). If yours isn't better, you're not done. Step 2: Split the job across specialist sub-agents instead of one overloaded agent. Step 3: Loop each piece until it looks like real work, not 'pretty good for AI'. Version 1 will be garbage. Step 4: Blind critics compare your output to a reference and keep going until they pick yours (or you stop). Why the critic must be blind: In Claude Code / Codex-style harnesses, sub-agents often fork the main agent's context, so a 'critic' still remembers the builder's rationale and rubber-stamps it. Fix: spawn critics with totally fresh context. They only see the artifacts (e.g. two images), pick which looks better, and don't know which is yours. Until the critic picks yours, keep looping. How to run: Lead agent decomposes, each piece gets builder + critic, before/after per subsystem, fold back into the game, repeat waves. Do not hand-feed a big pile of reference images; set the prompt and let the agent find comparison material itself. References when the thing doesn't exist yet: Use a different game/object as a quality-level comp (critic judges 'which looks better overall', not identity match), or generate target stills with an image model until you like them, then feed those as the bar. Same loop outside games: Writing: after each iteration, A/B paragraphs (or page/chapter) against strong recent comps. Avoid famous dead authors the model already 'knows' as themselves (IP + identity). Use contemporary comparable writing. Websites: feed sites you think are world-class; don't stop until blind critics consistently prefer yours. How prescriptive on 'what good means': optional. Experimental runs let loose. Real work: get more prescriptive and steer mid-loop ('like this direction, but pare back'). Cost and when to stop: The loop can run forever. Demos/toys: don't max it. Real work that matters: willing to spend a couple hundred dollars / hit subscription limits. Stop when the artifact is valuable and 'already wow', not because the bar was fully beaten.