Code-based generative art has always leaned on a kind of bricolage — artists combining well-known algorithms (Perlin noise, flow fields, L-systems, cellular automata) into new compositions rather than inventing new math from scratch. A new paper, “Wide but Shallow: Coding Agents in Generative Art” by Dominikus Baur, accepted at xCoAx 2026 (the 14th Conference on Computation, Communication, Aesthetics & X, in Torino), asks what happens when an LLM-based coding agent sits inside that bricolage process instead of a human doing all the assembly by hand.
The method: one challenge, two practices
Baur used the 31-day Genuary art challenge — a daily generative-art prompt format with an established community — as the testbed, running an autoethnographic study: comparing a manual coding approach against one assisted by coding agents across the course of the month, using the author’s own practice as the data.
A “difficulty collapse”
The paper’s central finding is what it calls a difficulty collapse: because generative art already leans on combining existing, well-documented algorithms, and because a coding agent can draw on a training set saturated with exactly those algorithms, “nearly any well-known algorithm can be used directly” without the artist needing to implement it from first principles. The barrier that used to separate an idea from a working sketch — actually writing the flow-field or L-system code — mostly disappears.
That collapse cuts two ways. It enables easier free-form experimentation and a wider variety of outputs, since trying five different algorithmic approaches to a prompt costs little more than trying one. But it comes “at the cost of less depth” — less time spent understanding why an algorithm behaves the way it does, less of the slow tinkering that used to be where a lot of generative art’s actual craft lived.
Where the paper thinks this goes
Baur frames the likely trajectory as a slow progression in artists’ roles: from inventing algorithms, to adapting existing ones, to — in the weaker version of the trend — simply changing parameters of algorithms an agent already wrote. The paper is careful to flag this as a possible “artistic surrender to the results from large-language models” rather than a settled verdict, but the pattern it documents (wide, fast, shallow) is a specific and testable claim rather than a vague worry about AI and creativity in general.