MIT Research: "Context Rot" Is a Harness Problem, Not a Model Limit

MIT CSAIL says agents trained on short contexts can generalize 8-32x longer, if the harness is built right.