One identical brief. Five CLI coding agents from different providers, each given the same prompt in YOLO mode and asked to build its own version — then peer-reviewed by a different model. A study in how model "personality" shows up in code.
Part of 0labs.dev · orchestrated with CLI Agent Orchestrator.
index.html
(inline CSS+JS, no build, no network) with an interactive canvas particle system. On load,
particles settle into the agent's own name; on mouse-move they scatter, then re-form.
Pick a color identity. Each agent also wrote first-person notes; a different agent reviewed it.
"One unified force model gives you the fly-in, the scatter, and the smooth re-form without any mode-switching logic."
"Code quality is solid and readable… one improvement: debounce the resize handler."
"The browser draws the word into an offscreen canvas, samples the visible pixels, and uses those as spring targets."
"The framerate-independent physics is a detail many demos skip… improvement: pre-render the halo and blit it."
"I kept the artifact brutally simple to run: one HTML file, no imports, no servers, no assets."
"Balancing polish with simplicity… movement remains readable rather than dissolving into noise."
"Electric cyan / periwinkle violet… a palette that reads as AI-coded without the lazy green-on-black terminal cliché."
"Solid, polished work… but the notes claim 'no comments in the code' while the source is heavily commented — a contradiction."
| Agent | Provider | Result | Personality in the code |
|---|---|---|---|
| CLAUDE | Anthropic Opus 4.8 | ✅ first-try | Most engineering-reflective; smallest, tightest file |
| HERMES | Hermes → gpt-5.5 | ✅ cleanest | Concise, almost poetic; "brutally simple to run" |
| CODEX | OpenAI gpt-5.5 | ✅ verbose | Most thorough physics; largest demo; terse notes |
| OPENCODE | minimax-m3 | ✅ slowest | By far the most-documented; its review caught it contradicting its own notes |
| KIMI | Kimi | ⛔ quota | Billing-cycle limit reached (403) |