What actually makes you trust a local coding agent enough to leave it running unattended?

I keep seeing people here saying they can leave local coding agents running for an hour or two and come back to something usable. I’m not really asking which model is best. I’m more interested in what made you trust the setup enough to stop watching it. For me, getting a task right once wouldn’t be enough. I’d want to see how it behaves when a tool fails, when the context gets long, when it hits an ambiguous requirement, or when its original plan turns out to be wrong. Tests help, but an agent can still pass tests while changing something you never intended it to touch. I think there’s a point where you stop trusting the model itself and start trusting the harness around it like permissions, checkpoints, git, tests, rollback, tool restrictions, etc. For people actually running local coding agents unattended, i just want to ask that, what was the thing that made you comfortable enough to stop babysitting them?