Evaluator-Optimizer
Generate, critique, refine — in a loop.
Evaluator-optimizer wires two LLMs into a loop: a generator produces a draft, an evaluator grades it against explicit criteria, and the feedback flows back for another pass. It's the machine version of a writer revising against an editor's notes.
The animation runs the loop live. The generator's first draft flows right to the evaluator, which stamps it — not good enough — and returns it with red annotations marking specific problems. The generator revises; the draft flows back; fewer red marks this time. After two or three cycles the evaluator fires its green ACCEPT stamp, breaking the loop and releasing the output.
Watch two things. First, the draft visibly improves each pass — that improvement is the pattern's entire justification. Use this loop only when clear evaluation criteria exist and iteration measurably helps; if the first draft is already at ceiling quality, every loop after it is pure cost. Second, the loop counter in the corner: iteration is bounded. A budget caps the cycles so an unsatisfiable evaluator can't spin forever — the loop ends on ACCEPT or on budget, whichever comes first.
One pattern remains: what happens when the loop's next step isn't fixed at all.