AI Agents Ship Dead Code, Bloat in Chrome Extensions
Manual review of AI-built TubeScribe extension uncovered dead code path, unneeded host_permissions, and 15KB bloat—fixes halved package size from 27.1KB and altered install prompt.
Overlooked Waste in AI Agent Output
AI agents building Chrome extensions like TubeScribe (YouTube transcript exporter to Markdown) produce functional first drafts but miss critical inefficiencies. Key issues: (1) dead code paths shipped to production, (2) unnecessary host_permissions scopes expanding user install prompts beyond needs, (3) ~15KB of bloat inflating the zip from essentials. Local upload was 27,766 bytes (27.1KB), but Chrome Web Store listed 31.83KB—highlighting store-side discrepancies plus agent flaws.
Agents prioritize working code over optimization, ignoring bundle analysis or permission minimization, leading to bloated deploys that erode trust via overreaching prompts.
Fixes Yield Measurable Gains
Removing dead code and excess permissions streamlined the install dialog, reducing asked-for access and improving user conversion. Stripping 15KB bloat cut package size roughly in half, from 27.1KB baseline—proving manual audits unlock 50%+ efficiency without rewriting core logic.
This second-pass review turns 'good enough' AI output into lean, production-ready code, directly tying supervision to faster loads and better UX.
Supervising Agents Requires Targeted Checks
Agents miss these because they emulate tutorials or docs without runtime profiling or security audits. Builders must enforce: bundle analyzers for dead weight, permission linters for MVPs, and code coverage to catch unused paths. First versions work but bloat; second passes ship—treat agents as drafters, not deployers.