Just to give a sense of my usual flow: In a private Slack channel I’ll @ my agent with a task and a few pointers. It’ll research internally (Slack, Google, Private GitHub, Jira) for evidence and if things line up it’ll open a PR. I’ll review the PR and leave comments and my Agent gets a notification when comments come into its PRs. Once the PR is merged the agent will follow along with the deploy and get notifications on each step. If needed it’ll update the Jira, send a customer a ping in Slack to let them know it’s merged and deploying and for them to test it out, and if things go wrong and need to roll back based on alerts or logs it’ll escalate to me.
Individually I could approve each step and sandbox it but that would limit how many of these things I can get going in parallel. Things have been OK for at least 6 months now with this flow. I suspect I could scale up to even more parallel agents if I could get this flow off of my machine but I dont think I’d ever go back to approving each of the steps. It would have to be on a new machine primarily to get a new identity which is slightly less scary but again no way I’m approving things step by step again.
It’s definitely scary to start with but once you are cooking it’s great.