Descriptions:
This video from All About AI explores a creative agentic experiment: using Claude Code, paired with Chrome DevTools Protocol (CDP) browser automation, to iteratively learn how to draw inside JS Paint by comparing its output against a reference image. The setup is intentionally minimal — no pre-built skills, no prior knowledge of the paint tool — forcing Claude to discover how to control the canvas from scratch using only CDP mouse control and screenshot feedback.
Inspired by a clip arguing that evolutionary ‘vibe coding’ (trying random mutations against automated test criteria) is the only path to complexity that would be impossible to hand-code, the host frames the experiment as a small-scale demonstration of that principle. Claude receives a goal image (a fisherman at a shoreline), navigates to JS Paint in Chrome, builds a CDP drawing script, renders an attempt, takes a screenshot, compares it to the reference, and loops. Across several iterations the output visibly improves — colors align, basic shapes emerge, text gets corrected when letters render backwards.
The experiment surfaces concrete observations about where current agentic tool-use excels (color fills, coarse geometry, iterative text correction) and where it struggles (fine detail, proportions). A second test with a hand-drawn ‘AI agent’ text image reinforces those findings. For developers interested in computer-use agents, visual feedback loops, or Claude Code’s capabilities beyond pure code generation, this is a practical reference point.
📺 Source: All About AI · Published March 16, 2026
🏷️ Format: Hands On Build
