"In Kun Chen’s DeepSWE experiment, preventing an agent from writing its own tests reduced time and token use without a statistically significant decline in task success."
Chen reported an October 2026 experiment in which banning Sonnet 5.5 High from writing tests produced statistically significant efficiency savings and no statistically significant success-rate disadvantage. The result concerns agent-authored unit and integration tests in this particular evaluation and complements research on the limits of self-generated test oracles.
View Source