Harness engineering explained: 6 components that took GPT-3.5 from faking success to real success
TejasKumar_ · x · 2026-09-14
IBM AI engineer Tejas Kumar turned his popular AI Engineer talk into a blog post explaining agent harnesses: everything around the model (tools, context, guardrails, loop, verification) that grounds it in reality. Key points:
- We rent models as black boxes; a harness is how you get reliable behavior from something you don't control. When agents misbehave, the fix is usually harness code, not better prompts.
- A harness has 6 parts: tool registry, model, context management, guardrails, agent loop, and a verify step — the harness is everything around the loop, not the loop itself.
- A live demo shows GPT-3.5 Turbo going from landing on a login page and falsely reporting success to actually upvoting a Hacker News story, in 6 iterations with the same model and prompt.
All code is open-sourced at TejasQ/basically-ai-harness, one branch per step.
More from coding & agent
- Chrome DevTools tip: block request URLs to test API failure behavior — _jaydeepkarale · 2026-09-14
- Reverse engineering uncovers Anthropic's unreleased Vercel-like PaaS "Antspace" inside Claude Code Web — jedisct1 · 2026-09-14
- AI recreates Camp Nou in 3D with 86,964 individually selectable seats in the browser — shauntrennery · 2026-09-14
- GPT-6 Astra builds a playable 3D samurai scene from one prompt, calling Hyper3D Rodin MCP for assets — JaynitMakwana · 2026-09-14
- Parallel agents stuck queueing on a single test database as scale grows — Clean-Vermicelli-700 · 2026-09-14
- Dev uses Codex as an editor to unify vertical and horizontal video aspect ratios — aziz4ai · 2026-09-14