Image-generation agents stop merely retrying: failure experience passes a counterfactual check before entering the skill library, reaching 86.1 on GenEval2

The scaffold around an image generator becomes a runtime that keeps accumulating repair experience under causal gatekeeping, instead of a control flow that resets after each task. Prevailing agentic generation pipelines remain shallow generate-verify-rewrite loops that treat vision-language-model judgments as coarse feedback, either discarding repair experience or admitting ungated updates that poison the skill library. On the Qwen-Image-2512 backend VEvo-Harness reaches 86.1 on GenEval2, 14.5 points above GenRouter; average scores are 76.28 on Z-Image-Turbo and 77.59 on Qwen-Image-2512, the best averages on both backends.

A failure no longer yields only a pass-or-fail boolean but is expanded into locatable root causes, and repairs are routed to different intervention types by root cause. Visual critics give only coarse binary feedback that cannot say where spatial distortion or attribute misbinding occurred, and monolithic prompt rewriting attributes every failure to prompt misalignment. The hierarchical verification tree expands sub-questions lazily only on failure and evaluates all sub-questions of all failed constraints concurrently, bounding diagnosis to one parallel vision-language-model round; in ablation the full setting is simultaneously the most accurate at 84.38 average, the cheapest at 6.2 calls and the fastest at 38.4 seconds.

Before a skill patch enters the skill library, counterfactual dual-arm replay estimates its causal effect, and only patches with positive estimated effect are admitted. Updating the harness by text-level success rates risks harness-level hallucination, where textually plausible edits degrade visual fidelity; without causal gating, admitted harmful patches run at roughly 33 to 40 percent within 100 episodes. Full visual continual governance brings the harmful-admission rate to about 2 percent by 100 episodes; removing the denied-change buffer or success preservation raises it to 10 percent and 8 percent respectively.

Persistent memory is organized into working, experience and pattern tiers, and recurring failure modes are distilled into preventive strategies injected before rendering. Flat episodic buffers transfer poorly and append-only skill registries degrade; removing persistent memory drops GenEval2 and KRIS by 6.7 percent and 6.2 percent. Over streamed tasks the rolling-average GenEval2 score rises from 63.2 to 86.5; without visual continual governance gains collapse after a mid-stream peak near 78 to about 71.5.

The demonstrated setting is an agent scaffold around a frozen generator: on the Z-Image-Turbo and Qwen-Image-2512 backends, with Qwen3.5-27B as the reasoning backend and at most 5 harness iterations, evaluated on five general generation benchmarks, four downstream tasks and four editing benchmarks.

The authors attribute the collapse of gains without causal gating to unverified updates polluting the skill library, an explanation drawn from admission-rate curves and ablation contrasts rather than an independent causal experiment. Cross-backend gains are reported under identical configuration, but evaluation stays within the listed benchmarks and backend combinations.