Dev runs MCP browser agent on real sites overnight, documents the bugs unit tests miss
JeremyCMorgan · x · 2026-11-18
Developer Trung Thanh Tran left Claude Code running with his turbo-browser-mcp attached to his own Chrome overnight (Nov 16), reading HN/dev.to/Reddit every 30 minutes, filling forms, creating a Reddit profile, and booking a Notion Calendar call. Six rounds of fixes became v0.7.0.
Bugs unit tests wouldn't catch:
- Two MCP servers on one Chrome hung forever: the CDP keep-alive bridge mixed up client sessions; the second client never received Runtime.executionContextCreated. Fix: isolated CDP sessions per client over one upstream socket
- One frozen tab could block every call: added a 1.5s cap on per-tab setup scripts
- Real-site failures: date fields silently clearing, placeholder selections passing validation, submission guards missing localized buttons
A practical failure-case collection for browser-agent builders.
More from coding & agent
- Musk says Grok Bot cut SpaceXSI invoicing from 24 hours to minutes with 5 accounting agents — elonmusk · 2026-11-18
- Users slam OpenAI's dots for lacking custom MCP server support — carlosdponx · 2026-11-18
- User runs 83 AI agents concurrently, nearly 300 combined — gnukeith · 2026-11-18
- Developer flexes running nearly 300 AI agents at once: "This is normal, right?" — gnukeith · 2026-11-18
- Token Maxxing Is Just Expensive Procrastination, Argues Viral Reddit Rant — Charming_Cod_2910 · 2026-11-18
- Team's hands-on with MCP Apps and Events: widgets work well, event push lags behind — lemuridaelabs · 2026-11-18