Last week I created two fairly complex programs with AI Coding Agents only and they worked as intended. That's hard to admit for me, given that I still can't wait for this whole bubble to pop and for the Silicon Valley to become a smouldering crater.
But while the programs technically work, it's crazy to see the agents handle errors. There is no logic or strategy in it: Throw every shell command in existence at it: ls, grep, cat and a little javac for good measure. It always seems to walk the solutions depth-first and try to make the wrong assumption work somehow.
It's literally this fantastic analogy of dumping rocks into the river until they can be used as a bridge.