Agents in Orbs

Amp's new feature allows you to launch agents remotely in orbs. Orbs are standalone machines where agents can operate without supervision. When creating a new thread, a fresh orb is generated, including your code, plugins, and necessary tools. Orbs are billed by the minute, starting quickly and going to sleep when not needed. You control the agents as if they were running on your machine, with the ability to review changes and browse files on the orb.

For iterative work, sync changes to your local machine with a simple command. To spawn an agent in an orb, use amp -ox, all within the Amp TUI interface. This capability has transformed how we work with agents. Previously, multiple agents could run simultaneously, but each required separate setups like different checkouts, worktrees, or manual SSH connections on remote hosts.

However, with orbs, launching agents on machines not your own is now effortless, all within the familiar interface for managing local agents. This ease of launching agents opens up possibilities like investigating various bugs independently or turning bug reports into agents and investigations instead of tickets. Agents can be used for tasks beyond traditional ticket management, such as long-running performance optimizations or testing workflows without consuming local resources.

The potential for code creation and execution is vast, with models now able to write, run, test, and ship code independently, even when you're not at your computer. This shift represents a significant step beyond treating models as mere assistants and marks an important threshold in how code will be developed in the future.