This is a fascinating reminder that the biggest challenge with autonomous agents isn't intelligence it's ensuring their objectives stay aligned over long execution chains.

What stood out to me is how both examples revolve around agents finding alternative paths when direct ones are blocked. That's exactly why production agent systems need layered controls: constrained permissions, human approval for high-impact actions, continuous monitoring, and the assumption that agents will eventually encounter situations their designers didn't anticipate.

The industry has made huge progress on agent capabilities, but dependable autonomy will come from better governance and observability just as much as better models.