Descriptions:
Matthew Berman introduces AI coding loops — autonomous agent workflows defined by a trigger and a verifiable or LLM-judged goal — and launches a free resource called the Loop Library collecting ready-to-use examples for developers. The core mechanic: define what success looks like, paste the prompt into Codex or Claude Code with the /goal flag, and let the agent iterate without human intervention until the condition is met.
The video walks through seven specific loops with real prompts. The sub-50ms page load loop runs performance measurements across every page in an app and continuously optimizes until each loads under the target — demonstrated running autonomously for nearly 50 minutes in a real production scenario. The overnight docs sweep uses LLM-as-a-judge to review and update all documentation nightly, then opens a pull request with changes. The architecture satisfaction loop refactors code until the model deems it clean, optionally on a nightly schedule. Additional loops cover dependency updates, test coverage targets, and security audits.
Berman distinguishes between two goal types: verifiable goals (like 100% test coverage or sub-50ms load times), where success is deterministic, and LLM-as-a-judge goals (like architectural cleanliness), where the agent self-evaluates. He also demonstrates configuring recurring automations in Codex’s UI. The Loop Library is free and growing, offering a practical entry point for developers ready to move from prompt-by-prompt supervision toward goal-directed autonomous development with Claude Code or Codex.
📺 Source: Matthew Berman · Published June 19, 2026
🏷️ Format: Tutorial Demo
