Self-Directed AI Dev: Cellular Automata (kenforthewin.github.io)

A benchmark tested several agentic LLMs in a simple autonomous development loop: a “Tech Lead” agent researched a coding task and spawned dev sub-agents (each with a chosen LLM and toolset) to implement a falling-sand cellular automata web page using only vanilla HTML/CSS/JavaScript. The spec: sand spawns at the top center, falls and piles at the bottom, and left-clicking the page spawns sand at the cursor. Runs were measured on cost, execution time, and lines of code using the Matic platform; results spanned about $0.06–$1.47, 2–12 minutes of wall time, and 87–1,142 lines of code. GLM 4.6 emerged the clear winner — both fastest and one of the cheapest — while some models produced much longer, slower outputs (one run yielded 1,142 LOC and higher cost). This matters because it shows measurable differentiation between agentic LLMs on practical engineering tasks: not just correctness but cost-efficiency, latency, and output conciseness in an end-to-end autonomous dev loop. The benchmark methodology—Tech Lead plus dev sub-agents—captures common real-world orchestration patterns for AI-assisted development, and the author plans to scale to more complex projects to probe models’ planning, decomposition, testing, and tool use. For practitioners, the takeaway is that model choice materially affects total cost and development velocity when deploying agentic coding workflows.