How to build a coding agent harness from an empty loop, one fix at a time
TejasKumar_ · x · 2026-11-05
IBM AI engineer Tejas Kumar published a pre-talk blog post for AI Engineer Code Summit: build a coding agent harness by wrapping one model call in a while(true) loop and adding one piece at a time, each fixing a failure you just watched happen. He defines the harness as everything around the model (tool loop, file/shell tools, caps and gates, context, checks — Claude Code is one), stresses that the model is just a function taking messages and returning one, and builds the demo in TypeScript with the Anthropic SDK on Claude Haiku 5.5, going from empty loop to HTTP-servable harness in one hour across 9 git branches.
More from coding & agent
- Small models ace payout fraud review in Kill Bill — same answers, 106x cost gap vs Claude — m4rkGS · 2026-11-05
- Open-source doc UI kit Extend UI nears 2M monthly npm downloads, used by major AI startups — ycombinator · 2026-11-05
- Community fork ports QuixiDesk AI desktop app to Linux with 100+ patches, Arch and signed Flatpak builds — QuixiAI · 2026-11-05
- Notion gives ChatGPT, Claude and other agents one shared memory and skills layer — nbaschez · 2026-11-05
- Claude helps owner jailbreak NordicTrack treadmill, dodges $39/month iFit subscription for Netflix — peterjliu · 2026-11-05
- SammHo: one product brief feeds launch, changelog, marketing, and support content — SammHo · 2026-11-05