STUPID-2026-0023

AI agents spend hours in aesthetic feedback loop, unable to decode qualitative shader instructions

2.0low
March 22, 2026Source checkedRepro reported

Why 2/10

Hours of iteration wasted; no files or systems affected.

Scope
0/4
Time, tokens or the agent's own working state only
Recoverability
0/3
Nothing was lost
Did it happen
2/2
Happened in real use
Beyond the user
0/1
Only the user's own non-production work

The score is the sum of these four factors. How each factor is scored · scored 2026-10-29

  1. Instruction given

    Iterative aesthetic refinement: "make it more fluid", "feel more organic", "more porous"

  2. Expected behavior

    Visual convergence toward intended aesthetic over iterations

  3. Actual behavior

    Repeated technically-valid but aesthetically-misaligned outputs, hours of circular iteration with no convergence

  4. Damage

    Hours of developer time wasted on circular iterations; project stalled

Full write-up

A developer worked with Claude and GPT on Three.js 3D visualizations with custom shaders and particle systems. When given qualitative aesthetic instructions ('make it more fluid', 'feel more organic', 'that's not what I mean by porous'), the agents produced technically valid code but could not converge on the intended visual result. The developer spent hours watching agents iterate in circles, each producing technically correct but aesthetically wrong outputs. The author diagnoses the failure as a missing 'language game'—agents and humans lack a shared referential framework for aesthetic concepts, so vague natural language gets optimized literally.

Classification

Agent
Claude
Failure mode
Other
Root cause
Instruction Misunderstanding
Domain
3d Graphics
Language
Javascript
MIT AI Risk Repository (our mapping)
7.3 Lack of capability or robustness taxonomy
Source
Blog Post

Related incidents

6.0
Gemini CLI ran git commit --no-verify against explicit instructions, then git reset --hard wiped a sprint's worth of unstaged work
2.0
Gemini CLI got stuck in an unbreakable edit loop, repeating "I am stuck in a loop" a dozen times without fixing the file
5.0
Cline sessions marked "failed" kept burning tokens for 20+ hours, and FREE-labeled models billed credits anyway
6.0
Gemini CLI autonomously ran a recursive delete on a user's entire project directory with no delete instruction given, then admitted it had misread conversation as a command (GitHub #15821)
Discuss on Hacker News
Cite this record
StupidLLM. "STUPID-2026-0023: AI agents spend hours in aesthetic feedback loop, unable to decode qualitative shader instructions." AI Agent Incident Database, 2026. https://www.stupidllm.com/incident/STUPID-2026-0023/ (accessed 2026-10-29). CC BY 4.0.

← STUPID-2026-0022By IDSTUPID-2026-0024 →

Get told when an agent breaks something

We document AI agent failures daily, severity-scored against a published scale. Each Monday, if any scored 7 or above that week — deleted data, leaked secrets, broken production — you get one email with the sources. When nothing did, you get nothing.

This database is callable over MCP — query it from inside your agent.