AI coding agents build frontend code they can never see, with no standardized browser-screenshot and console-error feedback loop

AI coding agents (Claude Code, Cursor, Lovable, Bolt) write frontend code but have no eyes: they cannot see the rendered output, detect layout breakages, or catch runtime console errors without a human running the browser and copy-pasting feedback. The HN thread on uncomfortable truths about AI coding agents (80 pts, 532 comments, March 2026) identifies 'the agent writes code but never sees what it built in the browser' as a core reliability failure mode. ProofShot (161+106 pts, March 2026) validated this gap with a working product, but it is a one-off CLI with no integration into agent loops. There is no standardized MCP tool or SDK plugin that gives coding agents a programmable browser headless loop with screenshot capture, console log collection, and structured feedback back to the agent context. This is a fresh gap specific to the agentic coding platform (2024-2026 era), not covered in the DB.