Building an Agent Harness From Scratch: GPT-3.5 Turbo Goes From False 'Done' to Real One, Prompt Untouched
tejaskumarlol · reddit · 2026-09-14
The full write-up (with code and links) behind the AI Engineer Europe talk: GPT-3.5 Turbo plus Playwright tools, tasked with upvoting the top Hacker News story, with the prompt identical across all versions — only the harness evolves:
- v0 bare loop: hits the login wall, still reports success.
- v1 guardrails: code-enforced iteration/message limits and naive context trimming.
- v2 verification: a deterministic function reads the tool call history to pass/fail each attempt, up to 3 retries — now it fails honestly.
- v3 harness-side login: the harness detects the login page and injects credentials the model never sees; done in 6 iterations.
Working definition: an agent harness is everything around the model that gives it grounding in reality — tool registry, context management, guardrails, the loop, and the verify step. Full blog post and per-step branches included; term sources: Mitchell Hashimoto's "harness engineering" and LangChain's anatomy of an agent harness.
Related event: IBM Engineer Explains Agent Harness: Making GPT-3.5 Actually Deliver(3 posts)→
More from coding & agent
- Creator builds shipwreck diving game with Opus 5.5 and Astrocade MCP — _amirabs · 2026-10-31
- GitHub Universe talk: 'Build once, run on any agent' with slides and skill files — 0xkarasy · 2026-10-31
- Claude can build pro software for ~$20k in tokens — and TAM shrinks to match — bigblueboo · 2026-10-31
- Microsoft launches Copilot harness in Foundry Agent Service for managed agent loops — WirelessLife · 2026-10-31
- Codex Remote live activities go live in the ChatGPT iOS app for tracking agent tasks — SIGKITTEN · 2026-10-31
- Open-source pixel FX bring a 3D wizard battle to life with slimes, archers and a golem — majidmanzarpour · 2026-10-31