I set up a separate account on my home server for a dev agent. It’s a separate account with limited permissions and the tools it needs installed. I also created a separate GitHub account for the agent, added it as a co-author for my repository, and configured Claude Code and Codex to work fully autonomously, without asking permission to perform any actions.
It’s a very different development experience. With the right set of instructions, an agent can implement a feature on its own from a fairly short description: create a branch, write a plan, code, and tests, create a PR on GitHub, read the comments on the PR, and make corrections. It’s all quite autonomous.
Yes, you still need to keep an eye on it and check what it does, but you’re not sitting in front of the screen waiting for another permission request. Of course, you can also run Claude Code or Codex in fully autonomous mode under your everyday work account, but you need to understand that all your data, passwords, and keys will be accessible to them and could easily leak. I wouldn’t take that risk.