Your agent didn’t fail when it hallucinated. It failed when it started being helpful.

You can hear agreement and still see misalignment.

In interviews, I learned to watch what people do instead of what they say.

Someone says they’re comfortable with the scope.

But their jaw tightens when autonomy comes up. Their breath pauses when you describe the timeline.

These are human micro-signals — small behavioral cues that reveal misalignment before someone says it out loud.

Agents produce signals like this too.

Not emotional signals — behavioral ones.

A recommendation appears that wasn’t requested.

A workflow continues one step past the assigned task.

An assumption slips in as “helpful context.”

I call this agent micro-drift — when behavior quietly expands the job.

Most teams evaluate agents by what they say.

But drift often starts earlier.