Screenshots Made a 4B Browser Agent Worse: A Postmortem With Real Numbers
Stunning-Sherbet1853 · reddit · 2026-11-11
Running a 4B model on a 6GB RTX 3050 laptop to compare prices across three shopping sites, the author found that adding annotated screenshots every step hurt: correct cells fell from 4/9 to 3/9, and 5/9 runs exceeded 25 minutes. Log digging revealed two self-inflicted bugs: image data counted as text (context overflow), and mismatched button-id instructions ("e12" vs real "j12") that guaranteed wrong clicks. After fixes, stuck-loop runs dropped from 4 to 1, but accuracy fell to 1/9 as step counts ballooned to 35–52, timing out 3 runs. The author's hunch: reading screenshots is itself a burden for a 4B model, and the critical step (picking the right product) doesn't improve. He asks the community for evidence of small models actually benefiting from screenshots.
More from coding & agent
- Developer removes Figma, Linear and VS Code from his dock as work goes headless — mobileraj · 2026-11-12
- Open-source music-skill lets Claude compose note by note — 7 pieces, 5 models, one prompt each — BandicootNo3152 · 2026-11-12
- MIT's Missing Semester adds an agentic coding lecture in its 2026 edition — moutsea · 2026-11-12
- DEATH MUST DIE: a game built with MCP and Claude Code officially launches — andrey_kurenkov · 2026-11-12
- Google Antigravity ships voice mode, cloud remote agents and /swarm in rapid-fire updates — gerardsans · 2026-11-12
- Stanford HAI's Loyal Agents initiative opens hackathon on human-agent research collaboration — StanfordHAI · 2026-11-12