Harness, Loop, Graph: why better prompts won't fix your broken AI agent
2C_ornot2C · x · 2026-11-08
@0xwhrrari publishes a long-form piece arguing that when your agent passes every test yet the issue stays open, a better prompt won't help — the broken part isn't the model but the system around it. The article breaks agent engineering into three nested layers: Harness (the shell wrapping the model: tool calls, context, constraints), Loop (the iteration cycle governing how the agent retries and verifies), and Graph (task orchestration controlling multi-step dependencies), explaining what each owns and which to fix first. Built and demonstrated in Claude Code, Codex, and Grok Bot; a year ago most of the work was in the prompt, now it's in this architecture.
More from coding & agent
- Open-source AEGIS simulates credential-leak outbreaks; adding memory lifts containment coverage by 7 points — shabrisharakosch · 2026-11-08
- Dev spends a week using Claude Code to animate a music video about side-project addiction — pizzaisprettyneato · 2026-11-08
- Incident agent stores only engineer-verified fixes in Hindsight to prevent memory poisoning — LingonberryMoist9805 · 2026-11-08
- Where should AI memory end and deterministic computation begin? A Reddit architecture debate — Strange_Wait7520 · 2026-11-08
- MCP servers can silently bypass host confirmation prompts after Day 1 approval — with a fix — Independent-Sir3234 · 2026-11-08
- Founder launches Topped, a hosted MCP server for AI-drafted and scheduled short videos — Ill-Basket-6162 · 2026-11-08