Agentic Pelican on a Bicycle

An experiment in agentic iteration on Simon Willison's "pelican on a bicycle" LLM benchmark.

The Prompt

Generate an SVG of a pelican riding a bicycle

- Convert the .svg to .jpg using chrome devtools, then look at the .jpg using your vision
  capabilities.
- Improve the .svg based on what you see in the .jpg and what's still to improve.

- Keep iterating in this loop until you're satisfied with the generated svg.
- Keep the .jpg for every iteration along the way.

Requirements

• Agentic model with vision capabilities, and an agent harness to run the loop
• Chrome DevTools MCP server (for consistent SVG rasterization)

Results

Each folder contains the iteration sequence for a specific model:

• claude-opus-4.1/
• claude-sonnet-4.5/
• claude-haiku-4.5/
• gpt-5-medium/
• gpt-5-codex-medium/
• gemini-2.5-pro/

Full Write-up

Read the blog post at: robert-glaser.de/agentic-pelican-on-a-bicycle