Every agent we ship has an explicit answer to one question: which actions can it take unattended, and which require a human to approve first? Getting this wrong in either direction breaks the system — too cautious and it's not actually saving anyone time, too autonomous and one bad call costs trust you don't get back.
The line usually falls at irreversibility and cost. An agent that drafts a refund request for approval is low-risk; an agent that issues the refund directly needs a much higher confidence bar, or a policy-based limit (auto-approve under $50, escalate above it).
We build the escalation path as a first-class feature, not an afterthought: a queue a human actually checks, full context on why the agent is unsure, and one-click approve/reject. Systems that route uncertainty to a Slack message nobody reads aren't human-in-the-loop, they're just quietly broken.
The goal isn't zero human involvement. It's making sure the humans still involved are only looking at the 5% of cases that actually need a person — which is usually where the real ROI of an agent shows up.