Your agent gets worse the longer it runs. The failure modes with their real names, the techniques that work, and the test that tells context from model.
Somewhere around the fortieth turn, something changes. The agent re-reads a file it read an hour ago. It proposes again a fix you already rejected. It answers a narrow question with a summary of the entire task, as though it had forgotten which question you asked. Nothing crashed. Nothing was deployed. The model string is the same one that was working perfectly at turn five.
Most practitioners file this under "the model got worse today". Usually it did not. What changed is what is sitting in the window, and there is a discipline whose whole job is to manage that. It has a name, a vocabulary, a small body of research and a much larger body of vendor engineering notes, and almost nobody hands the vocabulary to the person who needs it, which is the person watching a run go bad in real time.