Everyone wants agents that can deploy, roll back, and flip feature flags. Almost nobody wants an agent doing those things unsupervised at 2am. The real design question isn't whether to put a human in the loop. It's where to put them, and how, without making the agent useless.

I work on database DevOps, where approval gates are a fact of life: risky changes pause for a human before they proceed. The same idea applied to agents is a harder problem than it looks. MCP's elicitation flow lets a server pause and ask the human a question mid-task, but the tuning is brutal. Ask on every step and operators switch the agent off inside a day. Ask too rarely and you have shipped an unsupervised production-breaker.

This talk covers the patterns I use for human-in-the-loop DevOps agents: when to gate on approval and when to let the agent proceed, how to surface enough context for a human to decide in five seconds, and how to give each tool a risk level that both the agent and the operator can read. I close with a live approval-gated deploy, end to end, including what happens when the human says no.