Addy Osmani’s “Loop Engineering” thread argues that coding agents may be moving past the old “prompt, wait, prompt again” routine and toward systems that keep finding work, splitting it up, checking results, and repeating until a goal is met. Osmani presents the idea as an early and somewhat uncertain direction for agent workflows, while warning that token costs can climb quickly and that quality control still needs a close eye.
The thread reduces the concept to a small set of building blocks: scheduled automations, worktrees, skills, plugins and connectors, and sub-agents, plus a separate memory layer that keeps state outside the conversation. Osmani claims that both Claude Code and Codex already ship those pieces, which makes the model less about a single tool and more about how the workflow gets designed.
He then walks through how those parts appear to fit together in real use, from recurring triage jobs to isolated branches and separate verifier agents. The thread draws a clear line between the older habit of manually driving every turn and a newer approach where the system itself keeps asking, checking, and routing work until something is finished.
The replies that follow focus on the rough edges rather than the novelty. Commenters raise familiar concerns about token budgets, observability, memory files, security, and exit conditions, while Osmani notes in follow-up posts that those trade-offs are real and that limits still matter, especially on paid plans.
For developers watching how agentic tooling is changing, the thread is a compact tour of both the promise and the friction. The full post goes deeper into the mechanics, the caveats, and the examples behind the idea.
Source: Addy Osmani on X
