GPT-3.5 Runs Real Browser Tasks via Hand-Built Agent Harness, No Prompt Changes

A developer showed GPT-3.5 Turbo driving Playwright to upvote Hacker News stories with a fixed prompt, achieving real task completion purely through a hand-written agent harness and code fixes. Full code and writeup have been published.

2026-09-14 ~ 2026-09-14 · 2 related posts

• Agent Harness Lessons: Every Fix for GPT-3.5 Turbo Was Code, Not a Prompt — tejaskumarlol · 2026-09-14
• Building an Agent Harness From Scratch: GPT-3.5 Turbo Goes From False 'Done' to Real One, Prompt Untouched — tejaskumarlol · 2026-09-14