Colouring the acrobats in AGENTS

Near the top of the N, two backslashes touch. The first is a standing acrobat’s leg. The second is his neighbour’s raised arm. The colour has to change right between them.

The prompt

o           o__ __o       o__ __o__/_   o          o   ____o__ __o____    o__ __o    
        <|>         /v     v\     <|    v       <|\        <|>   /   \   /   \    /v     v\   
        / \        />       <\    < >           / \\o      / \        \o/        />       <\  
      o/   \o    o/                |            \o/ v\     \o/         |        _\o____       
     <|__ __|>  <|       _\__o__   o__/_         |   <\     |         < >            \_\__o__ 
     /       \   \\          |     |            / \    \o  / \         |                   \  
   o/         \o   \         /    <o>           \o/     v\ \o/         o         \         /  
  /v           v\   o       o      |             |       <\ |         <|          o       o   
 />             <\  <\__ __/>     / \  _\o__/_  / \        < \        / \         <\__ __/>  

Here's an ASCII art of the word 'agents', where each letter is made up of a number of small men. Use HTML to colour each man a distinct colour then serve locally making no mistakes.

Tracing the figures

I started by counting heads. Each lowercase o marked one acrobat, giving me 34 to trace. The first head sat nine columns to the left of its <|> torso. I restored the nine leading spaces on that row, then traced each pose on the grid until every visible character belonged to exactly one figure.

A monospaced <pre> kept the spaces and line breaks in place. I gave each acrobat a CSS class with its own colour and used that class on each row where the figure appeared. In the N, the two backslashes fell into adjacent <span> elements.

<span class="n-top-left">/ \</span><span class="n-diagonal-upper">\o</span>

I served the HTML locally and checked it in the browser at desktop and phone widths. All 227 non-space characters lined up with the corrected drawing, and the browser rendered one distinct colour for each of the 34 acrobats.

The finished artwork

How often can Codex get it right?

Jim’s Manifold market asks for the first year an AI solves this puzzle at least half the time. I’d count this drawing as one successful run. To make a case for 2026, we still need to see how often Codex can repeat it. I used the uncoloured Acrobatic FIGlet font to check poses in this run. Read the market rules.

I’d try ten fresh Codex sessions with the same model, settings and tools. Start each in a clean workspace, paste the prompt once, and keep every result. Count a pass only if the grid is intact and every stroke of each figure has that figure’s own distinct colour. Five passes out of ten would meet the 50% threshold in that sample.

The trial prompt puts Jim’s original two sentences first so the nine leading spaces in the market’s drawing survive the paste. The instruction is 181 characters, within the market’s limit of 333.

View trial prompt