Overview
- Coverage this weekend, with major reports published on Saturday and Sunday, shows leaders such as Anthropic’s Boris Cherny saying he no longer writes prompts and now relies on agent 'loops' to coordinate work.
- Loops are recurring, automated workflows that keep AI agents working toward a goal without a human typing new prompts, and Google’s Addy Osmani describes five core parts: automations, worktrees, skills, plugins/connectors, and sub‑agents.
- Practitioners recommend splitting roles inside loops so one agent writes code and a separate agent reviews it, because models tend to be biased when assessing their own outputs.
- Engineers warn loops bring clear operational trade‑offs: running multiple agents raises token usage and costs, adds orchestration complexity, and requires explicit stopping logic and independent verification.
- Adoption is strongest in coding workflows and open‑source tooling now, and observers say loop design could become a key product differentiator as teams balance faster automation against higher expense and extra safety checks.