What surprised us most while testing AI agents

While testing AI agents over the past few days, one thing surprised us:

They don’t fail the way traditional software does.

They don’t crash.

They don’t throw errors.

They behave… incorrectly.

And often, it looks completely normal.

A system can:

• follow the wrong instruction

• ignore safeguards

• change behavior subtly

And you might not even notice it.

That makes debugging and trust much harder.

It feels less like testing software,

and more like evaluating behavior.

Curious if others have seen this too:

→ Do your systems fail visibly?

→ Or do they fail silently?

Would love to hear how people are thinking about this.