Summary
Sharbel A. makes the case that most people using Claude Code, CodeEx, Cursor, or Hermes are effectively acting as the agent themselves — manually prompting, testing, and correcting one step at a time. The solution he proposes is what he calls loop engineering: giving an agent a complete process rather than a single instruction, so it can work, self-verify, and only surface to the human when blocked or done.
The video breaks down a loop into four components: a trigger (manual, scheduled, or event-driven), an execution skill (a reusable procedure the agent follows), a goal-and-verifier pair (so the agent knows when it has succeeded), and an output-plus-memory layer (so each run leaves a log the next run can inherit). Four filter questions help determine whether a task is even worth automating into a loop at all.
The hands-on portion shows a real loop built inside Claude Code that monitors a Notion board, automatically downloads filmed footage, runs a custom YouTube editing skill, generates a low-resolution preview for human approval, and only renders the final 4K version after sign-off. The full 37-minute autonomous run — with a verifier that forces the agent to retry any failed checklist items — illustrates both the power and the practical shape of loop engineering in production.
📺 Source: Sharbel A. · Published August 12, 2026
🏷️ Format: Tutorial Demo
