What counts as "human oversight" for AI agents?

As of August 2, 2026, the EU AI Act's high risk provisions are enforceable, meaning agentic AI systems now legally require human oversight and conformity assessments. Fines for non compliance can hit 15 million euros or 3% of global revenue.

What's interesting is that "human oversight" isn't clearly defined in most implementations I've seen. Does it mean a human reviews every action before it happens? Only high risk actions? A log someone can audit after the fact?

I've seen a few teams building agent products wrestle with this exact call for every action type, not as a compliance checkbox, but because it genuinely changes how much you can trust the thing day to day.

Curious how others here are interpreting "oversight" in what you're building or using. Full approval gate, audit trail, spot checks, or something else?