Asynchronous Coding Agent Control Loop Research Note

Asynchronous Coding Agent Control Loop Research Note

An asynchronous coding agent accepts a repository task, prepares an isolated environment, forms a plan, changes code, runs tools, and returns an artifact later. The developer moves from continuous steering to staged specification, plan review, result review, and integration.

Google's current Jules getting-started guide describes GitHub authorization, repository and branch selection, a cloud VM, optional environment setup, plan generation, and code changes. Jules task documentation says every task has its own VM, logs, environment, and changes.

An IDE assistant keeps the developer inside a faster interaction loop. GitHub's current Copilot IDE documentation distinguishes ask, plan, and agent modes.

The durable choice is not autonomous versus manual. It is whether the task benefits from delayed delegated execution or immediate human steering.