Kenan Tang

The Agentic De-Evolution (2026)

Recent vision-language agentic systems faithfully follow human instructions, producing outputs with perceptually negligible errors. Yet, this artwork asks what happens when humans step away and errors accumulate. By leaving an AI agentic system to operate on its own over an extended time horizon, this artwork investigates the long-term degradation of autonomous generative models, exposing their core weakness caused by a cumulative distribution shift.

To explore the error accumulation without human intervention, the artwork uses a popular agentic image editing system (Nano Banana Pro, Google) to perform a sequence of repeated edits. The system consists of an editing agent and an evaluation agent. The editing agent is given a simple task: exactly replicate a starting image of a colorful fruit platter. The evaluation agent then verifies if the generated image meets this requirement. Once the evaluation agent approves the output, it is directly used as the next input, and the process repeats for a total of 300 steps.

In the early steps, the outputs appear to be faithful copies. Over time, however, small errors accumulate. The initial image gradually loses its color and structure. Around 300 steps, the image eventually collapses into black-and-white striped noise.

More interestingly, this degradation is not uniform across space and time. Various hierarchies emerge over the editing steps: The background degrades faster than the central objects, and different fruits degrade at different paces. At certain points, the system even attempts to remove visual noise from some of the fruits, but inevitably fails in the end. This spatially and temporally uneven pattern shows a subtle, implicit bias in what the model preserves versus what it neglects.

Despite the visual collapse of the image, the evaluation agent remains blindly confident. Even as the output degrades into pure noise, the evaluating agent keeps its approval, describing the highly degraded images as “exact,” “flawless,” and “identical.” Without human intervention, the agentic system is trapped in a hallucinated certainty.

This artwork uses a single editing task and one model to represent universal weaknesses across a wide range of recent image editing and evaluation agents. A broader investigation can be found in our paper in CVPR 2026 Workshop on Agentic AI for Visual Media.

The Agentic De-Evolution serves as a cautionary observation. It questions the assumption that agentic systems consistently improve themselves over time. By showing how tiny deviations accumulate into full collapse, the artwork invites viewers to reflect on the risks of delegating critical societal functions to such fragile agentic systems.